AIPhoundry · Product and business builder for founders

We put the Ph in Phoundry: forging your digital ideas into Physical reality.

Bring us a product idea. We design it, prototype it, certify it, and launch the business that sells it, on your domain and your Stripe account. Then we hand you the keys: every file, every account, every supplier.

Five phases from idea to Physical

Every phase ends with deliverables you can hold, open, and keep. You sign off before the next phase is billed, so you never pay for work you haven't seen.

We're our own first customer

Before we take your idea through the Phoundry, we take ours. This site runs on the same template every founder's site launches from, and our first product goes through the same five phases yours will.

See what each phase produces

We test the Phoundry by running our own product ideas through it: a handheld sports gauge, a wearable translator, a robot cell, a control panel configurator, a smart inventory shelf, data center control panels, a solar installer platform with its own monitoring gateway, a configurator for intelligent motor control centers, a foundation crack repair kit sold alongside its own repair crew, a chicken coop door that counts the flock in before it closes, a modular desk rack for a Mac mini, Raspberry Pis, ESP32 boards and power banks, and a family of robot arms built from Rockwell motion parts. Each one is different, so each one pushed the platform somewhere new. Each tab opens with the case that the product is worth building: who has the problem, what the product changes, its price against what buyers use today, and its margin. These are cases on paper, built from synthetic tests, until real customers try them. Below each case are the real files the platform produced for them along the way: concept sheets, CAD, drawings, schematics, code, assembly instructions, test reports, and a demo store on its own subdomain.

These are demo projects, not customer products, and none of them is for sale. The designs are drafts that no licensed engineer has signed: phase sign-offs came from a labelled test reviewer and payments were simulated. Every bench, factory and field test result here is synthetic, made up to exercise the process, unless its card is marked Measured or Real.

Pickleball net gauge

Demo project · synthetic test data

A handheld gauge that sits on the net's top tape and reads its height to 0.1 in with a time-of-flight sensor, then shows PASS or RAISE/LOWER. Flip it over and the other end is a wrench for the net's winder.

  • Reads the net's height with a multizone time-of-flight sensor that ignores the mesh, then says PASS or RAISE/LOWER.
  • Flip it over and it's a socket wrench for the net's winder.
  • The first product to run every phase, Spec to Handover, and the first demo store with a 3D model buyers can turn and pull apart.

Open the demo store

Why it's worth building (Case on paper · synthetic test results)

Who has the problem
Club, park-district and tournament staff set many nets to 36 in at the posts and 34 in at the center. Today they do it by eye, with a $7 string gauge, or with a $51.99 aluminum height stick. None of those reads out a number or turns the net's winder.
What it changes
It reads the height to 0.1 in and says PASS within ±0.25 in, and the other end tightens the winder. In the synthetic sun test its worst average error was +0.09 in, with no dropouts in full sun.
Price against today's answer
$99, against $7 for a string gauge and $51.99 for a height stick. It has to win on speed and on a number staff can write down, not on being the cheapest.
Margin
The unit cost is $50.88 with the real Xometry quote for the printed enclosure, which gives 49% gross and about 36% landed. That is just under the 50% floor, so the enclosure is the cost to bring down.
What would prove it
Time ten real courts with and without the gauge, and ask five club managers what they would pay for it.
  • Concept sheet
  • Exploded enclosure CAD
  • Mechanical drawing
  • Visual assembly instructions
  • Pilot build results
  • Product render
  • Storefront

Translator badge

Demo project · synthetic device data

A badge worn on the chest that translates a conversation both ways. Captions face the other person on an outward screen; the reply plays through a speaker or the wearer's earbuds. Speech recognition and translation run in the cloud.

  • Captions face the other person on an outward screen; the reply plays in the wearer's earbuds or through a speaker.
  • Speech recognition and translation run in the cloud on Cloudflare Workers AI. Real timing tests picked the models.
  • A device plus a service: 6 months of translation included, then a monthly plan.

Open the demo store

Why it's worth building (Case on paper · measured cloud, synthetic device)

Who has the problem
Front-desk, retail, healthcare and event staff serve people who don't share their language. Today they pass a phone back and forth or buy a handheld translator or translation earbuds at $200 to $350.
What it changes
It works hands-free and face to face. Captions reach the other person in about 0.83 s and a spoken reply in about 1.9 s, for about $0.006 a minute of cloud. Those three numbers were measured on Workers AI; the device results are synthetic.
Price against today's answer
$169 with 6 months of translation included, then $3.99 a month, against $200 to $350 for Pocketalk, Vasco or Timekettle. A typical user's year of translation costs about $22 in cloud.
Margin
About 37.5% landed with printed cases, and 43.7% once the cases are molded (a $10k to $15k mold).
What would prove it
Lend ten badges to one hotel front desk or clinic for two weeks, and count how often staff reach for the badge instead of a phone.
  • Concept sheet
  • Cloud speed test
  • Product render
  • Exploded CAD
  • Screen designs
  • Schematic
  • Assembly step
  • Bench results

Robotic palletizer cell

Demo project · synthetic test data, fictional sites

A guarded robot cell that stacks cases onto pallets: a Comau NJ-40-2.5 arm run from Rockwell GuardLogix safety controls, with two pallet stations so one fills while the other is swapped. It is the first product in our second category, industrial machinery.

  • The first product in our second category, industrial machinery: a guarded robot cell, not a gadget.
  • Safety designed in from the start: ISO 13855 safety distances, NFPA 79 schematics, and safety-rated PLC and HMI programs.
  • Sold by quote request, with a finance-lease option, instead of a buy button.

Open the quote-request store

Why it's worth building (Case on paper · assumed costs, synthetic tests)

Who has the problem
Stacking cases by hand at the end of a packing line takes about 2 people a shift. Over 2 shifts that is 4 people at about $55k each, roughly $220k a year (assumed figures), all doing heavy, repetitive lifting.
What it changes
Rated at 80% of the cell's computed capacity: 21 light cases a minute (two per pick) or 10.5 heavy ones. Two pallet stations let one pallet be swapped while the other fills; with heavy cases, the forklift's swap time sets the real rate.
Price against today's answer
$310,000 a cell (assumed), or about $6,954 a month on a 60-month finance lease, against about $220k a year of hand stacking. On those assumptions it pays for itself on labor in about 17 months.
Margin
33% gross and 25.3% landed, against a 25% floor. That is tight: a 10-point rise in tariffs on the robot alone would take it under.
What would prove it
Get real quotes for the robot and guarding, and time the palletizing labor on one real packing line.
  • Concept layout
  • 3D cell model
  • Layout with safety distances
  • Electrical schematics
  • Factory acceptance test
  • Build and install instructions

Control panel builder

Demo project · synthetic orders and tests

Built-to-order industrial control panels, UL 508A style. A buyer describes the machine (motors, I/O, voltage) and a configurator produces the whole build package for a panel shop to follow.

  • One request in, a full build package out: drawings, a 3D model, CNC code for the door and backplate cutouts, the bill of materials, the wire list, and wire cut lengths.
  • Ladder schematics drawn from the panel's own netlist, plus wire markers, labels and a manual.
  • Orders outside the rules are refused automatically instead of being built wrong. This is the first product in our third category, control panels.

Open the demo store

Why it's worth building (Case on paper · real packages, synthetic orders)

Who has the problem
Plant engineers and machine builders need a few control panels a year and have no panel shop of their own. Before anyone can build or even price a custom panel, someone has to draw its drawings, bill of materials and wire list.
What it changes
One request goes in and a complete build package comes out, with no hand edits. All 10 pilot packages were generated with 0 check errors (real, computed). Three bad requests were refused before anything was built: a door handle mounted too high, a 600 V supply and a 100 kA fault level.
Price against today's answer
$12,800 to $26,200 a panel in the pilot, priced by a fixed rule from parts and build hours, with a 15-working-day lead time. What buyers pay today for drawings plus a one-off shop build has not been benchmarked yet.
Margin
32% gross on every pilot panel and 26.8% landed overall, against a 25% floor.
What would prove it
Send three real machine requests to two local panel shops, and compare their quotes and lead times with the configurator's.
  • Enclosure and door machining
  • Backplate layout
  • Configurator runs
  • Power schematic
  • Ten pilot orders
  • Product render, door open
  • Product render, door closed

VMI inventory tracker

Demo project · synthetic device data

A shelf that counts its own stock. Four lane scales weigh the bins in a vendor-managed inventory area, and the supplier and the customer both see the same live counts, so reorders go out before anything runs short.

  • Counts parts by weight in each of four lanes, with an e-paper label on every lane.
  • Its own LTE-M cellular eSIM, so it never joins the customer's network.
  • A shared dashboard where the supplier and the customer see the same stock, and reorders raise themselves. You can try a live demo of it.

Open the demo store · Try the live dashboard

Why it's worth building (Case on paper · synthetic bench and demand data)

Who has the problem
In vendor-managed inventory a supplier keeps a customer's shelves stocked. Today a rep walks the shelves once a week with a scanner, so stock-outs happen between visits and usage is only known afterwards.
What it changes
It counts every bin all the time. On the bench model it counted 1,000 picks and restocks exactly. Over 180 simulated days its own reorders kept stock-outs about level with the weekly walk (7 days against 6, out of 3,600), with 10% less stock, 73% fewer pieces short, and no weekly visit.
Price against today's answer
$449 a unit (about $112 a bin) plus $12 a month, against a rep's weekly visit to every customer site. The cost of that visit has not been measured yet.
Margin
The hardware costs $186.76 to make: about 58% gross and 44.2% landed. The $12 monthly plan costs under $0.10 a unit to run.
What would prove it
Put five units at one real supplier's VMI site for a quarter, and compare stock-outs, stock carried and rep hours with the quarter before.
  • Concept sheet
  • Counting core
  • E-paper label screens
  • Supplier dashboard
  • Assembly instructions
  • Pilot build and field run
  • Product render
  • Exploded view

Data center control panels

Demo project · synthetic orders, simulations and tests

Control panels for data center power and cooling, built on Rockwell parts: a power management master panel that sheds load and starts generators, a monitoring panel, and a liquid-cooling pump panel.

  • Every panel is rated 100 kA short-circuit current at 480 V, checked by a calculator that follows UL 508A Supplement SB.
  • Rockwell parts with a redundant ControlLogix PLC application and HMI screens generated for each order.
  • The sequence of operations is proven in a simulated 2N power plant before any panel is built.
  • Three products from one configurator, which reuses the control panel builder's tools.

Open the demo store

Why it's worth building (Case on paper · simulated sequences, synthetic orders)

Who has the problem
US data center construction ran at about $85B a year in August 2026, up from $49B a year earlier (founder's sourced research). The gear, and the people who wire and commission it, are the bottleneck, and the control panels around the switchgear are often left to the electrical contractor to find.
What it changes
Every panel is rated 100 kA at 480 V as standard, with its PLC application generated for each order. The power plant's sequence passed 7 of 7 simulated failure cases, including losing both utilities, generator trips and a controller switchover, and it caught two deliberately broken versions of the logic.
Price against today's answer
$245,000 for the power management master, $72,000 for monitoring and $52,000 for liquid cooling, with a 50-working-day lead time. It competes with integrators' one-off panels and switchgear makers' bundles on flexibility and lead time, not on price.
Margin
34% to 40% gross depending on the panel. The ten-panel synthetic pilot came to $991,000 at a 33.3% landed margin.
What would prove it
Bid one real contractor's data center project, and see whether the generated submittal package and the lead time win it.
  • Concept sheet
  • One-line diagram
  • Sequence of operations, simulated
  • Short-circuit rating
  • Backplate layout
  • 3D model
  • PLC application
  • Ten pilot orders

SunLoop solar installer platform

Demo project · synthetic device data

One system for a residential solar installer, from the first lead to twenty years of monitoring: incentives with their sources, quotes, jobs, parts orders, and a gateway beside the main panel that reads any major inverter and battery brand and compares each home's output with what the quote promised.

  • Quotes that show every incentive with its source and the date it was checked, and never price one that needs review.
  • A job pipeline with gates: the 120% panel rule, parts ordered, California disclosures and the cancel window, and a commissioning check that catches reversed CTs before inspection.
  • A monitoring gateway that reads SolarEdge, SMA, Fronius, Sol-Ark, Generac and Enphase, and alerts the installer when a home falls below its promise. You can try the installer app.

Open the demo store · Try the installer app

Why it's worth building (Case on paper · synthetic homes and installers)

Who has the problem
Small residential solar installers, with 5 to 50 staff, keep six tools in step: design, CRM, incentives, parts, and a separate monitoring app for each brand. The federal homeowner credit ended for systems placed in service after 2025, so they have to cut costs and sell more batteries.
What it changes
Quotes catch problems before they go out: 3 of 5 sample homes failed the 120% panel rule once their battery was counted. In the synthetic pilot (100 gateways, 4 installers), commissioning caught 11 install errors before inspection, 6 of them reversed CTs, and the promised-versus-actual alert found about 412 kWh of lost production before any homeowner called.
Price against today's answer
$499 a month per installer, plus $349 a gateway and $5 a site a month, in place of separate tools for design, sales, incentives, parts and each brand's monitoring.
Margin
The gateway costs $144.90 to make: about 58% gross and 40.4% landed at $349. The monthly platform and monitoring fees come on top.
What would prove it
Hold ten installer interviews, then sign three design-partner installers for the three free months in the pilot offer.
  • Brand coverage and incentives
  • Gateway core and quote engine
  • Installer pipeline
  • Promised vs actual
  • Gateway LED states
  • Work instructions
  • Pilot build and field run
  • Product render
  • Exploded view

BucketWise MCC configurator

Demo project · synthetic build and test data

A configurator for intelligent motor control centers in the CENTERLINE 2100 style. You list the loads; it lays out the lineup, sizes the bus, plans the network and writes the PLC program, and the maker builds and lists the lineup under UL 845 while the founder programs, tests and commissions it.

  • Packs starters, feeders and drives into sections by space factor, from small in-unit drives to full sections and 40 inch drive bays up to 500 hp.
  • Sizes the horizontal and vertical bus, picks the bracing up to 100 kA, and checks phase spacing, support spacing against the fault force, heat per section, harmonics and arc energy reduction.
  • Every motor is networked over EtherNet/IP and keeps a manual override: Hand-Off-Auto or Local-Off-Remote, and drive bypass, that still works with the network or the controller down.
  • One configurator makes three lineup sizes, and refuses anything out of range (a 700 hp drive, for example) and sends it to an engineer.

Open the demo store

Why it's worth building (Case on paper · synthetic build and test data)

Who has the problem
EPC firms, water utilities and process plants buying a motor control center wait through weeks of back-and-forth for a budget layout. Then they pay separately for the PLC program, the network and the factory test.
What it changes
It lays out and prices an engineered, networked lineup in minutes, with the PLC program and an integrated factory test included. Its sequence tests caught all 31 deliberately broken versions of the logic, and it refused a 700 hp drive that is outside its range and sent it to an engineer.
Price against today's answer
$264k small, $799k mid and $1.826M large per lineup. A buyer can ask the same distributor for the bare MCC, so the edge has to be speed, the engineered program and the integrated test, not hardware price. That competition risk is still open.
Margin
30.4% gross at $799k. The three-lineup synthetic pilot came to $2.89M at a 26.8% landed margin. A split margin, lower on hardware and higher on engineering, is proposed.
What would prove it
Quote one real EPC's MCC schedule, and set it beside the distributor's own price for the bare MCC.
  • Concept lineup
  • Large lineup, generated
  • Bus layout and spacing
  • Override and network-loss check
  • Drive bypass schematic
  • Three pilot lineups
  • Integration instructions
  • Mid lineup render
  • Large lineup render

CrackShot crack repair

Demo project · synthetic test and field data

A leak repair for cracked basement walls, sold two ways from one store. Homeowners can buy a kit that injects water-activated polyurethane foam into a leaking crack from inside, or book a crew to do it with a lifetime transferable warranty. One set of assessment rules decides which crack fits the kit, which needs the crew, and which needs a structural engineer instead.

  • The platform's first chemical product: resin lots are released against the maker's certificate, every kit is traced to its lots and use-by date, and it ships ground only as a Limited Quantity hazardous material.
  • One store sells the kit by checkout and the crew service by quote, with a homeowner's form (town and ZIP, how many cracks) instead of a company's. The platform gained per-product sale modes to do this.
  • Assessment rules refuse horizontal, stair-step, bowing, offset and widening cracks and block walls, and send them to an engineer. They are written into the checker, the box, the instructions and the crew's process.
  • The pilot covered both halves: 100 kits packed and shipped, plus 25 paid crew jobs, each checked for leaks after the next rain.

Open the demo store

Why it's worth building (Case on paper · synthetic test and field data)

Who has the problem
Poured-concrete basement walls crack and let water in after rain or snowmelt. Waterproofing companies quote $500 to $1,000 a crack, and caulk or hydraulic cement often pops out within a season.
What it changes
In the synthetic pilot, 89 of 100 kit buyers were dry after rain with one kit, and 6 more needed a second kit for a long crack. The 25 crew jobs (36 cracks) had 1 callback, 2.8% of cracks, fixed on the return visit. The assessment rules sent 5 requests away from injection, 3 of them to a structural engineer.
Price against today's answer
The Leak Kit is $129 for about 3 m of crack. A crew repair is $695 for the first crack and $395 for each one after, with a lifetime transferable warranty. Today's answer is $500 to $1,000 a crack from a waterproofer.
Margin
The kit costs $46.67 to make, about 64% gross. A crew job averages $932, leaving about $376 (40%) after the crew, materials, leads and the warranty reserve.
What would prove it
Sell 20 real kits in one metro and book 5 real crew jobs, then check every crack after the next heavy rain.
  • Concept kit and injection
  • Bench trials
  • Hazard label and coverage table
  • Tested instructions
  • 100 kits and 25 crew jobs
  • Leak Kit render
  • Kit, front view

RoostCall coop door

Demo project · synthetic test and field data

An automatic pop-hole door for backyard flocks of 2 to 12 birds. Every door on sale closes on a timer or a light sensor, whether or not every hen is home. RoostCall counts the birds in and out with a small sensor over the doorway, closes when the whole flock is in, tells your phone when one is still out, and closes anyway at your curfew so the coop is never open all night.

  • Counts in the dark with an 8 × 8 time-of-flight sensor: no camera, no leg bands or tags. The counting code was proven on synthetic sensor frames, including birds side by side, birds that stop in the doorway and birds that jump through.
  • A doubt never closes the door early. If the count is uncertain, the door waits for the curfew, and the owner can correct the count with one tap.
  • A self-locking lead-screw drive that a raccoon can't lift, a sensor that confirms the door is down and locked, and stall detection that backs off, retries and sends an alert.
  • Runs on a 2 W solar panel and a LiFePO4 battery that won't charge below freezing. It keeps working without Wi-Fi, which is only used for alerts and the owner app.

Open the demo store · Try the owner app

Why it's worth building (Case on paper · synthetic test and field data)

Who has the problem
Small-flock owners lose hens that are still outside when a timer or light-sensor door closes at dusk, so many walk out every evening to count birds anyway. Doors on sale run from $40 to $250, and none of them knows whether every bird is in.
What it changes
In the synthetic pilot, 100 doors ran 6,000 dusks. 5,412 closed with every bird in. 339 closed at curfew with a bird out, and each of those sent an alert first. The audited counts were 98.5% right. Two doors closed with a bird out. A fix for two birds passing through together then took that to 0 in the regression.
Price against today's answer
$209 plus shipping, against $199 to $229 for the leading app-connected doors that close on light or a timer. It costs about the same as a smart door and is the only one that counts the flock.
Margin
The unit cost is $100.83, about 52% gross. With freight, duty, returns, warranty and payment fees it lands at about 36%, so the sheet metal and battery quotes are the first costs to bring down.
What would prove it
Fit 20 doors at real coops for a month, check each night's count against the owner's, and count the evenings no one had to walk out.
  • Concept
  • Counting bench
  • Owner app
  • Exploded view
  • 100 doors, 6,000 dusks
  • Store render
  • Inside the coop

StackSmith desk rack

Demo project · synthetic test data

An 8U desk tower on the 10-inch rack standard for the small computers and boards that pile up beside a monitor. A Power Spine at the back holds your own power strip and USB-C charger, so every cable is short. Slide-in modules hold a Mac mini, six Raspberry Pis, ten ESP32 dev boards and five power banks. It was the first product with no electronics of its own, so it needed a new category in the platform, hard goods, with its own fit, load, stability and transit tests.

  • Mac mini sled: four cradles hold the M4 8 mm above a perforated tray so it breathes from below, and a front cap presses the power button underneath through a push rod and bellcrank.
  • Raspberry Pis stand on edge as blades at 36 mm pitch, room for an Active Cooler with an M.2 or PoE+ HAT, and slide out with the board.
  • ESP32 dev boards from 17 to 31 mm wide grip on cards, ten to a 1U rail with a write-on label strip. Nothing else in the 10-inch world holds them.
  • Five power banks stand upright in open bays to charge, with their cables through rear notches. Any module also fits a DeskPi RackMate or other 10-inch rack.

Open the demo store

Why it's worth building (Case on paper · synthetic test data)

Who has the problem
A desk with a Mac mini, a few Pis, a drawer of ESP32 boards and some power banks needs four products from four vendors: a $120 10-inch frame, a $49 Mac mini mount, a $34 to $135 Pi case, and nothing at all for the ESP32 boards or the banks. None of them holds the power strip or the USB-C charger, and the M4's button and air intake are underneath.
What it changes
In the synthetic prototype tests, the Mac mini ran 2.6 °C warmer than on an open desk with no throttling, and 1,000 of 1,000 presses on the front cap turned it on. Six Pi 5 boards at full load stayed under 68 °C, and the loaded tower did not tip until 27.5°. In the synthetic pilot, 94 of 100 units passed inspection the first time, and 88% of buyers would recommend it.
Price against today's answer
$299 plus shipping for the full stack, against about $200 to $300 for a frame, a Mac mini mount and a Pi case bought separately, which still leaves out the ESP32 boards, the banks and the power.
Margin
The unit cost is $144.66, about 52% gross. With US-made steel, freight, fulfilment, payment fees and a warranty reserve it lands at about 41% with shipping charged. Sheet metal is about 70% of the cost, so the Xometry quote is the number to confirm first.
What would prove it
Put 20 units on real homelab desks for a month, with each owner's own Mac mini, Pis and banks, and log what didn't fit and how often the front button missed.
  • Concept
  • Fit survey and bench
  • Exploded view
  • Mac mini sled
  • 100 units
  • Store render
  • From the front

AxisForge robot arms

Demo project · synthetic test data

A six-axis arm, a SCARA and a delta built from Allen-Bradley servo motors and drives and bought gearbox units, for machine builders whose machines already run on Rockwell. There's no robot controller: the arm runs in the machine's own GuardLogix with Logix kinematics, on the same safety system and spares. It needed a new platform category, robotics, with sizing, stopping and safety-function checks.

  • Every joint is a bought strain wave or planetary unit with its own output bearing, so no made part has a bearing seat: only a pilot bore, a bolt circle and two dowels.
  • Links are stock aluminium tube clamped between plate cheeks. A different reach is a different tube length, and every made part can be quoted as CNC plate, tube or MJF, with no castings and no molding.
  • A 6 mm zero pin through each joint masters the arm without a laser tracker; multi-turn absolute encoders keep it through power cycles.
  • Kinetix VP on Kinetix 5700 by default, or Kinetix TLP on Kinetix 5300 where it saves cost and space: the delta ships on TLP, 22% less for arm and drives.

Open the demo store

Why it's worth building (Case on paper · synthetic test data)

Who has the problem
When a Rockwell machine needs a robot, the builder buys one from a robot brand, and with it a second controller, a second programming language, a second safety system and a second spares crib. The robot's controller and pendant alone run $25,000 to $60,000 with a 5 kg arm.
What it changes
In synthetic prototype tests the A6 repeated to ±0.033 mm, the S4 to ±0.017 mm and the D3 to ±0.07 mm, and every safety function reached PL d on demo device data. The first stopping test found the A6's J1 coasting 61° with no brake, so it got one. In the synthetic pilot, 7 of 10 arms passed first time and the rest after rework.
Price against today's answer
$51,900 for the A6 with motors, plus about $13,000 of Kinetix drives in a machine that already has a GuardLogix. An Asian 5 kg arm with its controller costs less on its own; AxisForge wins where it removes the second controller, safety system and spares crib.
Margin
A6 unit cost $36,006, 30.6% gross, about 26% landed with freight billed at cost. About 80% of the cost is bought motors and gearbox units, so distributor pricing moves the margin more than the frame does.
What would prove it
Build one A6 for real, run it in a customer's Rockwell machine for a month with an integrator teaching it, and measure the stopping data and the teach time against a robot brand's arm.
  • Concept
  • Kinematics and joint bench
  • Exploded view
  • Assembly drawing
  • Three test arms
  • 10 arms
  • The family

TeachPair desktop robot arms

Demo project · synthetic test data

An assembled and calibrated leader and follower pair of the open-source SO-101 robot arm, designed by TheRobotStudio with Hugging Face (Apache-2.0), for people who teach robots by demonstration with LeRobot. TeachPair keeps the design and credits it; it changes only what building it in volume needs. It needed a new desktop class in the robotics category, with FCC, listed-supply and age-grading checks in place of the industrial robot standards.

  • Built on someone else's open design: the platform warns until the license, the credit and a list of every change ship with the product, and never presents it as the founder's own.
  • The FDM parts are hollowed for MJF nylon with a 2.5 mm skin and powder escape holes, and every servo pocket is opened 0.14 mm, by scripts that only take material away.
  • The assembly is checked on upstream's own STEP files, placed where upstream's robot model puts them: no clashes, and every overlap is one of upstream's own press fits.
  • Each arm ships with its servo IDs set, a factory torque limit and cut-off, and its own LeRobot calibration file.
  • Each arm has its own controller, a Seeed bus servo board with a XIAO ESP32C6: LeRobot works unchanged over USB, and the leader can drive the follower and replay recordings without a computer (firmware written and compiled, not yet run on hardware).

Open the demo store · The upstream SO-101

Why it's worth building (Case on paper · synthetic test data)

Who has the problem
A lab that wants to record robot demonstrations can print and build an SO-101 pair in a weekend for about $300 of servos and prints, or buy an assembled FDM pair for about $500. Either way it sets servo IDs, calibrates and debugs before the first episode.
What it changes
In synthetic tests the followers repeated to ±0.8 to ±1.0 mm against a ±2 mm spec, held the rated 0.08 kg straight out, and passed 200 hours of replayed teleop and two shipping drops with calibration intact. In the synthetic pilot, 8 of 10 units passed first time.
Price against today's answer
$899 for a calibrated Pair, about 80% over assembled FDM pairs: for labs and schools that pay for a warranty, support and data on day one. A Follower on its own is $399.
Margin
Pair unit cost about $674 with the MJF prints at JLC3DP's real rate ($0.377 per cm³ from the Desk S4 quote, about 1.7 times the formula the spec used): the Pair's 821 cm³ of prints cost about $309. That was 10% gross at $749, so the Pair went to $899, about 25% gross. The Lab Set and Follower are still near 10 to 14%.
What would prove it
Quote the print set at JLC3DP, build three Pairs for real, and put them in a university lab for a term against an FDM pair, counting setup hours and failures.
  • Concept
  • Hollowing and sizing
  • Exploded view
  • Assembly drawing
  • Three test Pairs
  • 10 units
  • The Lab Set
  • Desk S4 SCARA
  • Desk D3 delta

PointRail distributed I/O

Demo project · synthetic test data

Low-cost distributed I/O for Rockwell CompactLogix and ControlLogix. An expander head on a Raspberry Pi CM5 shows up in Studio 5000 as one Generic Ethernet Module and runs an EtherCAT chain of 32-input, 16-output and 8-channel analog bricks behind it. It needed a new industrial controls class in the electronics category, with Class A FCC, UL 61010-2-201 and EtherCAT and ODVA conformance in place of consumer checks. PointRail is a working name.

  • One head carries about 2,100 points at a typical mix: the POC computed what fits one 500 / 496-byte generic connection (the spec had guessed 4,000), so bigger systems add heads, one PLC node each.
  • Studio 5000 ready: the head's EDS file and a generated L5X with a UDT per brick, a named tag per channel and a copy routine.
  • Reading OpENer's source showed its default 10 ms timer would turn a 2 ms RPI into 10 ms and its 512-byte buffer can't hold a 500-byte image: the head app sets a 1 ms tick and a 1,500-byte buffer.
  • Four boards drafted and placed (not yet routed), housings in MJF nylon with a clean geometry check, and the bricks' safety firmware (100 ms watchdog, safe state, CRC-checked chip reads) passing its host tests.
  • An end-of-line station tests every unit on a known-good chain under a real CompactLogix connection, and writing the supplier list caught a 10/100 jack on the 1 Gb plant ports.

Open the demo store

Why it's worth building (Case on paper · synthetic test data)

Who has the problem
A panel with 2,000 I/O points on 5069 Compact I/O runs about $36 a point in list-price modules (estimate), and a big system can run out of the controller's Ethernet nodes.
What it changes
One PointRail head is one node for about 2,100 points. In the synthetic POC the CM5 held a 1 ms EtherCAT cycle with 94 bricks for 24 hours with no lost frames, and a 5069-L320ER held the full-size connection at 2 ms for 8 hours with no faults. In the synthetic pilot 94 of 100 modules passed end-of-line first time.
Price against today's answer
$8.24 a point for the 2,000-point reference system, about a quarter of Compact I/O's list price, or $1,127 for a starter kit to try it.
Margin
Kit parts $388.25 on the BOM r4 estimate, 65.6% gross at $1,127; the 2,000-point system's parts come to $6,678, 59% gross. Two prices are web-checked; the rest are estimates until real quotes.
What would prove it
Route the boards, build the head image and run a real CM5 and CompactLogix bench, then put a 500-point system in an integrator's test panel beside Compact I/O.
  • Concept
  • Points per connection
  • Exploded view
  • Assembly drawing
  • 100 modules
  • Starter kit

CartWright mobile robot

Demo project · synthetic test data

A light-cart AMR for labs, print farms and staff-only areas that carries 100 kg or tows a 200 kg cart and docks itself. It aims for open-source-robot cost with the navigation decisions of a big warehouse AMR: it maps, localizes, avoids people and re-routes with ROS 2 Nav2. It needed a new mobile class in the robotics category, with mobile-robot safety standards (ANSI/A3 R15.08, UL 3100), battery shipping and proof of navigation at every phase.

  • The platform refuses a robot for floors shared with untrained people unless it has a certified safety scanner. So the $7,500 Lab version is sold only for controlled spaces, and every channel says so: the store, the labels, the manual and the first step of commissioning.
  • Navigation comes from ROS 2 Nav2: slam_toolbox mapping, AMCL, an MPPI controller and a docking server, on a Jetson Orin Nano with a 360° lidar, a depth camera, four corner ToF sensors and two cliff sensors.
  • The camera and ToF layers keep only what the lidar and the saved map can't explain, such as pallet-jack forks under the lidar, so shelves and people don't box the robot in.
  • E-stops and bumpers run through a hardwired dual-channel safety relay. If the motor driver fails, two contactors short the motor windings to brake the robot, and the manual gives stopping distances for both cases.
  • It has no molded parts: a 40-series aluminium extrusion frame, cut and bent sheet, and MJF nylon corner caps and pods.

Open the demo store

Why it's worth building (Case on paper · synthetic test data)

Who has the problem
Labs, print farms and small shops that move totes between benches either push carts by hand or look at warehouse AMRs. Those carry certified safety scanners and are sold by quote at factory prices. A ROS research base costs less, but ships without docking, safety functions or support.
What it changes
In a simulation built to stand in for Nav2 (not Nav2 itself), the robot ran 50 of 50 missions in a sample warehouse with people, glass and low obstacles: 5 of 5 docks and no collisions. With the lidar alone it hit something 13 times. In synthetic prototype tests it docked 150 times out of 150 and ran over 12 hours a charge. In the synthetic pilot, 8 of 10 robots passed first time.
Price against today's answer
$7,500 with a dock, a spare kit and half a day of commissioning, for controlled spaces only. Buyers in the synthetic pilot compared it with AMRs at $25k and up.
Margin
Built cost $3,126 on the spec r3 BOM, 58% gross at $7,500. With freight, the crate, the spare kit, commissioning, a warranty reserve and insurance, the landed cost is about $4,287, 43% gross. Most prices are estimates: quotes are still owed.
What would prove it
Run the same missions on real Nav2 in Gazebo. Then build three robots and run them in a university lab and a print farm for a term, counting missions, interventions and stops.
  • Concept
  • Navigation in simulation
  • Exploded view
  • Assembly drawing
  • Three test robots
  • 10 robots
  • In use

AisleKit build plans

Demo project · synthetic test data

A plan generator for DIY builds: the buyer picks a project, and AisleKit writes the whole plan from a 3D model, with the shopping list (store model numbers, item ids and prices), a cut list nested on the board lengths and sheets the store sells, 1:1 printable templates, step-by-step drawings and a fit check. The buyer pays for the plan and buys the parts at the store. It needed a new plans category on the platform, where the product is a download, and the buyer's parts are the buyer's cost, not the maker's.

  • Nothing in a plan is typed by hand. A checker reads the finished PDFs back and compares every length, listing, count, step and template hole with the model. The CS-1 plan passes 369 of 369 checks, and when errors were seeded into the documents on purpose, it caught 7 of 7.
  • The cut list is nested on what the store sells, and the shopping list buys the cheapest mix that covers it. Each line names an alternative or a substitute in case the store is out.
  • Made to fit: change the bays, rows or deck-box size and the whole plan is made again, checked, and drawn again in about a minute.
  • Prices go stale, so a weekly refresh re-prices each plan. A plan whose total moves more than 10% waits for a reviewer, and a plan whose part is gone is held, never sold with a guessed price. A buyer's download stays the exact file they bought.
  • The same generator wrote three demos: the CS-1 card display case, a 4-hen chicken coop, and a utility sink hookup whose drain is checked against plumbing code figures.

Open the demo store

Why it's worth building (Case on paper · synthetic test data)

Who has the problem
Builders find plans online, mostly $2 to $15 on Etsy or free on blogs, and then work out the parts themselves. The plan rarely matches the board lengths and sheet sizes the store sells, so they buy too much, make a second trip, or cut a piece short.
What it changes
Every AisleKit plan is generated and checked from one model, with a shopping list for one store's shelves and a cut list nested on what it sells. In the synthetic build test, three first-time builders finished the case in 3 h 55 min to 5 h 25 min, on receipts of $104 to $109 for the required parts.
Price against today's answer
$14 for a plan, or from $24 made to fit, against $2 to $15 for a plan that leaves the parts to the builder.
Margin
About $1.54 landed a plan with the payment fee (89% at $14). The parts are the buyer's, paid at the store. The data-feed and support costs are estimates until the feed is licensed and real buyers write in.
What would prove it
Get a licensed product feed, then have real first-time builders build from the plan and compare their receipts and times with the plan's.
  • CS-1 card display case
  • The same generator, a coop
  • Plan pages
  • Three test builders
  • 100 plans and five price refreshes
  • Utility sink hookup

PhaseKit kits by phase

Demo project · synthetic test data

Software a lumber yard runs to deliver a house's framing in five kits, one per phase: deck, walls, deck, walls, roof. The estimator enters or imports the job once, and PhaseKit lays out the framing to the code edition the job names, picks each phase from the yard's own bays and lengths, nests the saw cuts, bundles everything in install order, and plans the truck and the drop. It grew out of AisleKit and needed a new kitting category on the platform, where the product is software and the lumber is the builder's cost. Its second trade, PhaseKit Workcell, runs on the same core: an automation distributor kits a machine builder's robot cell in six phases, from the builder's released BOM and configuration sheet, with every drive, controller and switch in the series and firmware the cell was engineered with.

  • Every kit is checked against the job model before it prints: each framed member in exactly one line of one kit, each cut equal to its member, each load within its truck, and every page read back from the PDF. On the demo plan that is 1,385 checks, and 9 of 9 errors seeded on purpose were caught.
  • Load plans stack bundles so the ones needed first come off first, count tiedowns per stack to 49 CFR 393.110(b), and put boxed hardware in the cab. Drop maps keep set-downs out of the overhead-line zone.
  • Imports a yard's item master and stock from BisTrack-style and Spruce-style exports, with a code map for the yard's own item numbers, and the truss plant's layout. PhaseKit never sizes trusses, LVL or I-joists.
  • A field change makes a change kit: the adds and the returns for each released phase, checked so nothing is counted twice.
  • Overage is set per item class from the yard's closed jobs, and every kit ends with a terms page: the accuracy, code, delivery and data notices.
  • Workcell matches the configuration, not just the part number: the builder's configuration sheet sets each configured part's series and firmware, stock is taken only from matching lots, older firmware gets a flash task at pack-out, and series and firmware are scanned and written on the ticket. Safety-rated and configured parts never get an alternate.
  • Workcell puts the whole cell on one order: the panel's terminal blocks, breakers, supplies and enclosure, the 80/20 extrusion cut at the distributor's saw, the pneumatics and the fasteners ship in the phase kits with the drives and controllers. A part the build is waiting for leaves the day it lands.

Open the demo store · Open the Workcell demo store

Why it's worth building (Case on paper · synthetic test data)

Who has the problem
A yard ships a house's framing in a few big drops from an estimator's takeoff. Material sits on site for weeks, the crew hunts for the right pieces, and short items mean extra trips the yard pays for.
What it changes
PhaseKit kits each phase from the yard's stock, cut and bundled in install order and checked against the job model. In a simulated pilot of 12 made-up houses at 2 made-up yards, short trips fell from an assumed 1.6 and 1.2 a job to 0.5 and 0.67, and estimator time from an assumed 4 h to about 1 h.
Price against today's answer
$149 a house kitted plus $490 a month per branch, against about $350 a house of takeoff time, trips and returns (an estimate the pilot is meant to measure).
Margin
About $14 a house to run (compute, storage, model check, support, all estimates), or 90.6% gross at $149.
What would prove it
Run one real yard's past jobs and compare with what it shipped, then time its estimator and load a real truck from the plan.
  • A house, phase by phase
  • Pick ticket by bay
  • Against a seller's material list
  • Load plan with tiedowns
  • Against past jobs
  • 12 houses at 2 yards
  • The demo plan's framing model
  • A robot cell, phase by phase
  • Release report with lead times and trucks
  • The right configuration, every cell
  • The whole BOM on one order

Phactory proof runs

Demo project · real POC in progress, nothing built yet

After 17 demo products, design is no longer the slow part. Every one of them stopped where the work met the real world: real carts, real builders, real tests. Phactory is the platform's execution layer. It runs a Proof Run, a small real batch of 3 to 10 units built from a locked design package by someone who has never seen the design, with every unit tested and logged and a named engineer signing the result. Its own proof of concept can't be synthetic: it builds 5 real translator badges, or it fails.

  • Source: every BOM line priced at each supplier with shipping, duty and stock, US and in-stock first, using the platform's buy-before-make engine.
  • Buy: lines grouped into as few carts as possible and batched across products. A person approves and places every order.
  • Make: a job ticket and a question log for an outside builder. Every question is a gap in the package and becomes a fix in the next revision.
  • Prove: a bench test per unit that writes a log tied to its serial number. Only measured results are filed as real evidence.
  • Sign: a named engineer signs the batch or rejects it with reasons. A new execution pack refuses anything but real evidence for these exit criteria.

Open the demo page

Why it's worth building (Case on paper · estimates)

Who has the problem
The platform takes an idea to a locked design and a demo store in hours, but proving it in the real world runs at human speed through one person. Across 17 demo products, no real unit has shipped, estimates were up to 7 times off until a real quote came back, and over 100 open conditions wait on real-world proof.
What it changes
Phactory turns a locked design into real units and real evidence: a landed-cost plan, one approved batch of orders, an outside builder, a test log per unit, and an engineer's signature, with what it learns written back into the next revision.
Price against today's answer
Demo prices: $3,500 a Proof Run plus parts, builder and engineer time at cost +10%. The first run's own pass-through is about $3,000.
Margin
About $2,950 of the $3,800 Phactory keeps on a 5-unit badge run, or 78% (estimate).
What would prove it
Build the 5 badges for real: at least 3 by someone other than the founder, every unit's test log filed, and the real landed cost within 20% of the plan.
  • The first Proof Run
  • The board in the run
  • Landed-cost plan for 5 badges
  • Firmware for the build
  • Order batch and job ticket
  • phactory.aiphoundry.com

TierLink power and data hub

Demo project · synthetic test data

An injector in the control cabinet puts the controller's Ethernet and up to 8 A at 60 V DC on one T1 hybrid cable. IP67 hubs on the machine serve IO-Link and Single Pair Ethernet devices, and every port identifies its device before it powers it. It was built from a draft cabling tier specification, as that draft's reference design. TierLink is a working name.

  • Identify before power on every port: detect, wake at about 100 mA, identify, grant from a budget the hub measures, and a firmware trip at 110 % of the grant for 100 ms.
  • Simulating the bench on the real port firmware found five design faults before any hardware, including a calibration that put a good 24.9 k signature outside its own window.
  • Two of five boards routed pin by pin with KiCad DRC clean (0 violations, 0 unconnected). Drawing them from the data sheets caught a 24-pin cable that could not carry two RMII ports and a power path that could not keep its input off below 30 V.
  • End-of-line test nests generated from the routed boards' own test points, with 13 tests and numeric limits.
  • A BOM with a supplier, lead time and alternate on every one of its 80 lines, and the five single-source chips named.

Open the demo store

Why it's worth building (Case on paper · synthetic test data)

Who has the problem
An IP67 IO-Link master on a machine needs an Ethernet cable and a separate 24 V power cable, and a powered field device that is plugged into the wrong port gets power before anyone knows what it is.
What it changes
One T1 cable carries both, and each port identifies its device before it powers it. In the synthetic pilot, 94 of 100 units passed end-of-line first time and 23 over-budget grant requests were refused as intended.
Price against today's answer
$795 for a hub with 4 IO-Link and 3 SPE-with-power ports, against about $385 to $1,500 for an 8-port IP67 IO-Link master with EtherNet/IP today (market research, list prices).
Margin
Starter kit parts $559.75 on the BOM r4 estimate, 56.3 % gross at $1,282. Nine of 80 prices are web-checked; the rest are estimates until real quotes.
What would prove it
Route the three port boards, then build the POC bench (about $3,150 of eval boards and cordsets) and run identify-before-power and 10BASE-T1L under 8 A load steps on real hardware.
  • Concept
  • Bench in simulation
  • Field board routed
  • Exploded view
  • 100 units
  • TL-H8M hub

FloorSlate thin-client tablet

Demo project · synthetic test data

A 10.1 in rugged tablet for the plant floor. Seated in its 24 V dock it boots Rockwell ThinManager firmware over the plant network; lifted out, it keeps the session on Wi-Fi. Two hot-swap battery packs, glove and wet touch and a badge reader under the glass. A browser kiosk mode runs web HMIs for plants without ThinManager. FloorSlate is a working name and is not ThinManager Ready.

  • Four boot modes: ThinManager over PXE in the dock, Windows IoT with the TMC app, a locked browser kiosk for Ignition, FactoryTalk Optix or WinCC Unified, and the ThinManager Ready BIOS, which needs a Rockwell partnership and is not offered.
  • Two hot-swap 37 Wh packs in the Adafruit style (pin 1 GND, a fuel gauge per pack). The embedded controller takes a pack off the rail before its contacts open, with 470 µF of hold-up computed for the changeover.
  • Dock board routed pin by pin, with KiCad DRC clean (0 violations, 0 unconnected). Its 14 spring pins carry no 15 V until a tablet's detect pin grounds them.
  • Embedded controller firmware for the RP2350 built in CI and checked by 53 host tests.
  • End-of-line station with 12 tests and numeric limits, including PXE boot into ThinManager, hot swap under load, a leak test and the BIOS profile. The dock's test nest was generated from the routed board.

Open the demo store

Why it's worth building (Case on paper · synthetic test data)

Who has the problem
Operators walk to a fixed thin client to see a line's screens. The rugged tablets that boot ThinManager are sold through Rockwell's hardware partners at prices that are only quoted.
What it changes
The tablet boots from the plant's own ThinManager server in any dock and keeps the session on Wi-Fi when lifted out. Plants without ThinManager can run it as a web HMI kiosk instead. In the synthetic pilot, 93 of 100 units passed end-of-line first time, and one site that refused Secure Boot off moved to the Windows mode.
Price against today's answer
$2,290 for the tablet with two packs. The ThinManager Ready tablets it competes with (Pepperl+Fuchs, Teguar and others) publish no prices, so there is no list price to compare.
Margin
Tablet parts $1,074.70 on the BOM estimate, 53.1 % gross at $2,290. One of 57 prices is web-checked; the rest are estimates until real quotes.
What would prove it
Build the real bench (about $1,450: the module's starter kit, the panel, the radio, the reader and ThinManager on its demo licence) and boot ThinManager firmware on it, docked and undocked.
  • Concept
  • Checks before the bench
  • Dock board routed
  • Exploded view
  • 100 units
  • Tablet in its dock

Pay phase by phase, not all at once

Hardware is expensive to guess at, so we don't ask you to. Each phase gets its own quote. You see the deliverables, sign off, and only then is the next phase billed. Stop whenever you like and keep everything you've paid for.

Submit your idea

Questions founders ask

What kinds of products can you build?
Any physical product is in scope. We're starting with electronics: connected devices, smart-home gear, anything with a circuit board and firmware. Other categories open up as each one proves itself in production.
Who owns the design?
You do. Your contract assigns you the design IP, and you are the legal manufacturer of your product. We design it; you own it.
Who manufactures my product?
Your manufacturer does, starting with the 100-unit pilot. You contract with them directly, and we hand them a revision-locked build package with assembly instructions, the test fixture design, and a quality plan. AIPhoundry designs, certifies, and launches your product; we don't manufacture it.
Can I leave partway through?
Yes, at any phase. You leave with every file, every account, your supplier contacts, and any tooling you paid for.
Is AI designing my product?
AI agents draft the designs, the code, and the copy, which is how we move fast. Design rules check every draft, and a named engineer reviews and signs off each phase before anything is ordered or built, unless you choose AI Engineer review.
What is AI Engineer review?
An optional, faster track. Instead of a named engineer, an AI reviewer checks your Spec, Prototype, and Pilot designs against the same design rules and checklists, usually in minutes, and you accept each review before anything is ordered. It costs 20% less on those phases. AI review can miss problems a person would catch, so boards may need respins and designs may fail certification; you take that decision. A named engineer still signs off Launch, accredited labs still test for certification, and it isn't offered for products with mains power, lithium batteries, or over 100 W. You can switch back to engineer review at any time.
I already have a prototype or some code. Do I start over?
No. Point us at your repo and we import what exists, then start at the phase you've actually reached. Our own first product did exactly that.
Who handles certification?
We scope the certifications your product needs, such as FCC, CE, and RoHS for electronics, and run them as part of launch. Certificates are issued in your name because you're the manufacturer. Nothing goes on public sale until its FCC authorization is confirmed.
How does selling work?
Your storefront launches on your own domain and your own Stripe account, so buyers pay you directly. Every paid order becomes a manufacturing order locked to the exact revision that buyer paid for, with no re-typing between checkout and factory.
How long does it take?
It depends on the product. Your spec quote includes a schedule for each phase, with the riskiest question tackled first.
What does it cost?
Spec review is $2,500, credited in full toward your prototype. After that credit, a typical connected product totals $52,000 in service fees across every phase, plus parts and lab fees at cost plus 10%. Each phase is a fixed quote, billed only after you sign off the one before.
Do you take equity?
No. You pay a fixed fee per phase. The optional Partner plan defers part of your fees against a capped share of sales, never equity and never a claim on your IP.

AI agents: see llms.txt to browse and order through the API.