Bench 1 · Optics & vision
Cameras, lighting, and the model that reads them. Line-scan, global-shutter, thermal. This is where a wet-glass glare problem gets solved with a polarizer before it ever gets solved with a network.
Applied R&D · 21-day burns · two partners
Light the work that should not wait for a committee.
Moonfire is a small lab. We take one problem — sensor fusion, edge inference, or industrial software that has to survive a shop floor — and we run it as a 21-day burn. On day 21 you get a working artifact and the source, not a roadmap. If it is still a slide deck, we failed and you say so.
21dstandard burn, fixed
3concurrent benches, hard cap
0equity taken unless you ask
2partners. That is the whole lab.
House rule 1. Ugly and working beats pretty and pending.
House rule 2. Source lands in your repo, not ours.
House rule 3. We test at 40°C or we do not ship it.
Live view: Cycle 14 — mill-vibration classifier for Kesler Stamping. Day 13 of 21. Enclosure sitting at 41°C on the test rig.
Burn remaining
8d
Cycle 14. Eight working days to a shippable classifier.
Bench load
Three benches, no fourth. A booked lab says no.
Ship-risk dial
Open question: does it still classify at 40°C after two hours? Until that is green, risk stays here.
Vibration signal, mill #3
Raw accelerometer trace before the model. The spike at frame 96 is a chipped tooth. That is the whole job.
Confidence split
71%
71% clean, 17% flagged, 12% "send a human." Moonfire tunes the middle band, not the accuracy vanity number.
21-day clock
Day 13. The clock is the contract. We do not extend it quietly.
What we actually build
Hardware-adjacent software that has to work in a loud room.
No pure-cloud SaaS, no chatbots, no "AI strategy." We take the messy edge — a sensor, a PLC, a camera, a machine that vibrates — and we make software that reads it and decides something in under a second.
Sensor fusion
Accelerometer + current draw + temperature into one signal. We fuse the cheap sensors you already have before we ask you to buy a nice one.
Edge inference
Models that run on a Jetson or a bolted-down mini-PC, not a data center. If it needs the internet to decide, it is not an edge model.
Industrial software
OPC-UA, Modbus, and the one weird serial protocol your 1998 machine speaks. We write the adapter and hand you the wire captures.
Latency-bound decisions
Reject-arm signals, go/no-go gates, anomaly holds. We budget the millisecond path and prove it on a scope, not a slide.
Field-survivable code
It runs after a power blip, a fat-fingered reboot, and a shift change. Watchdog, safe-state, and a log an operator can read.
Honest benchmarks
We report the number that breaks the demo: the 40°C run, the dirty-lens frame, the 3 a.m. shift. Vanity metrics get cut.
Jetson Orin
OpenCV / TensorRT
OPC-UA
Modbus TCP
Rust firmware
ROS 2
Grafana on-prem
RS-485
Three benches, no fourth
The whole lab is two people and three benches. Here is what each one is.
Bench 2 · Edge GPU
Where the model gets small enough to run on a bolted-down box. Quantize, prune, profile on target, budget the millisecond path. If it needs a data center to decide, it does not leave this bench.
Bench 3 · Firmware & wire
The unglamorous half: the adapter for the 1998 PLC, the watchdog, the safe-state, the log an operator can read at 3 a.m. Held this cycle for a part on backorder — and we say so.
We cap at three benches on purpose. A fourth means a burn goes shallow, and a shallow burn ships a demo instead of a machine. When all three are lit, the answer is a start date, not a rushed yes.
The 21-day burn
Five gates. Miss one and we stop the clock with you, not around you.
- Days 1–3 · FrameWe come to your floor, not a call. We define the one number that has to move and the failure that would make you fire us. Signed in a page.
- Days 4–8 · CaptureInstrument the machine, record real signal in real conditions. No synthetic data until we have the ugly stuff. You keep every raw capture.
- Days 9–14 · BuildModel and firmware together. First artifact runs on target hardware by day 14 or the burn is declared at risk and you get the call.
- Days 15–19 · Beat on itHeat, vibration, dirty inputs, power loss. We try to break it in your conditions and log every failure with a timestamp.
- Days 20–21 · Hand offSource in your repo, a run-it-yourself README, and a one-page verdict: what it does, what it will not do yet, and what a Line pilot would take.
Cycle 12 · actual timeline
Ravenna Bottling — fill-level vision check on a 900-bottle-per-hour line.
- Floor walk
Found the real problem was glare off wet glass, not the model. Reframed the burn in an hour.
- 4,200 frames
Captured across two shifts and one cleaning cycle. Kept the bad lighting on purpose.
- First on-target run
Ran on the line PC at 22 ms/frame. One day early. Rare.
- Broke it
Condensation on the lens dropped recall to 71%. Added a dirty-lens guard that pages the operator.
- Shipped
98.1% catch on underfills, source handed over, one open item logged. No day 22.
Day 21, in the box
What you actually walk away with. No mystery deliverables.
Handed to you
- The working artifact, running on the target hardware we agreed on.
- Source committed to your repository, not a private one of ours.
- Every raw capture and label from the burn, in an open format.
- A run-it-yourself README that a new engineer can follow cold.
- A one-page verdict: what it does, what it will not do yet, what a Line pilot adds.
- The failure log — every way we broke it and what happened.
Not in the box
- A per-seat license fee. You bought the work, you own the work.
- A dependency on a Moonfire cloud. It runs on your box, offline.
- A data-retention clause that keeps your captures on our disks.
- A roadmap deck instead of a thing that runs.
- An 18-month "phase 2" you have to sign to get phase 1.
The test we hold ourselves to: could you fire us on day 22 and keep shipping with what is in the box? If the answer is no, we did not finish the burn.
Bench log
The last six burns, with the number that mattered.
| Cycle | Client (sector) | Problem | Days | Result | Verdict |
|---|---|---|---|---|---|
| 09 | Halden Forge (metal) | Furnace anomaly from current draw | 21 | 94.6% | Shipped, Line pilot live |
| 10 | Corliss Dairy (food) | Cap-torque drift detection | 19 | 0.9 Nm | Shipped early |
| 11 | Vantek Tooling (CNC) | Tool-wear from spindle audio | 21 | 88.0% | Shipped with caveat |
| 12 | Ravenna Bottling (bev) | Fill-level vision on wet glass | 21 | 98.1% | Shipped |
| 13 | Nordic Ply (lumber) | Knot grading at 60 m/min belt | 21 | — | Failed. Refunded burn. |
| 14 | Kesler Stamping (metal) | Mill-vibration fault classifier | 13/21 | live | In burn now |
Cycle 13 failed. The belt was too fast for the camera we could source in 21 days, and we said so on day 16 instead of stretching to day 40. We refunded the burn and wrote the honest post-mortem. We would rather lose the fee than the name.
Ship rate, trailing
Q14/5
Q25/5
Q34/5
Q45/6
Heat-soak test grid
Each cell is a 10-minute interval on the thermal rig, floor to 44°C. Green stays green: no accuracy cliff.
Bench status
B1optics / visionRUNNINGc14
B2edge gpuQUEUEDc15
B3firmwareHOLD—
Bench 3 held for a part on backorder. We do not fake a busy lab.
Proof & boundaries
What operators say, and what we will not take.
“They showed up on the floor day one, found out our real problem was glare not software, and still shipped in 21 days. First lab that didn’t try to sell us a platform.”
— Plant manager, Ravenna Bottling (Cycle 12)“Cycle 13 didn’t work and they told us on day 16 and gave the money back. That’s why we booked Cycle 14 the next quarter.”
— Operations director, Kesler StampingWe take the burn when…
- There is a real machine or signal we can instrument this month.
- Someone on your side can point at the failure that would get us fired.
- A working-but-ugly artifact in 21 days is worth more than a study.
- You want the source and the raw data in your hands, not ours.
We pass when…
- It is a pure-cloud SaaS or a chatbot. Not our floor.
- You need a signed 18-month roadmap before day one.
- The only success metric is a demo for a board meeting.
- You want us to keep the data hostage to keep the contract.
Why a burn
The 21-day burn against the things operators usually settle for.
| Path | Time to a running thing | What you own on day 1 | Who keeps the data | Fails loudly? |
|---|---|---|---|---|
| Moonfire burn | 21 days | Working artifact + source | You | Yes — refunded |
| Big consultancy | 6–9 months | Slides + a SOW | Them | Rarely |
| University grant | 12–18 months | A paper | Shared / murky | No |
| Hire an ML team | 4–6 months | Salaries + ramp | You | Slowly |
| Off-the-shelf SaaS | Days | A tool that ignores your floor | Them | N/A |
We are not always the right answer. If the problem is a decade of instrument science, that is Deep Star. If it is a solved SaaS category, buy the SaaS. The burn is for the specific thing on your floor that nobody has made work yet.
Engage
Three ways in. All fixed-scope, all fixed-clock.
Scout
$8k
10 working days
- Floor visit + signal capture
- Feasibility memo with the real risks
- Go / no-go for a full burn
- No prototype — honest scoping only
Burn — most booked
$28k
21 calendar days
- Working artifact on target hardware
- Source committed to your repo
- Every raw capture is yours
- Day-21 verdict, no day 22
- Failed burn = refunded burn
Line
custom
~90 days
- Pilot hardened on your floor
- Runbook + operator training
- Handover to your team
- Equity? Only if you bring it up.
Moonfire is not Deep Star Labs. Deep Star is the long stare — multi-month instrument science on a research calendar. Moonfire is the 21-day burn: one problem, three weeks, a thing that runs. Different labs, different clocks. If you need the long stare, we will point you there.
What if you miss day 21?
You get the call by day 16 if we see it coming, not a surprise on day 22. A Burn that does not ship is refunded. We keep the reputation, you keep the cash.
Do you take equity?
Zero by default. We are a fee-for-work lab. If you would rather trade some fee for equity, we will talk — but you have to raise it, we never will.
Who owns the code and data?
You do, both. Source lands in your repository during the burn, and every raw capture is yours to keep even if we fail. We do not hold data to hold you.
Only three benches — what if you are full?
Then we say so and give you a start date. A lab that is always available is a lab that is not shipping. Three concurrent burns is the hard cap for two partners.
Can a Scout turn into a Burn?
Yes. The Scout memo is the first ten days of a Burn if you convert. You pay the difference, not a second kickoff.
Talk to the lab
Book a burn.
Two partners read every message. Tell us the machine, the signal, and the failure you are afraid of.