We turn factories into RL environments for physical AI.
Manifold deploys AI into industrial sites to run their operations. The work our AI does produces the outcome-graded training data physical AI companies can't get anywhere else. We sell both.
Two sides, both stuck.
The physical world is full of sensors and starved of intelligence. The AI world is full of intelligence and starved of physical data. Nobody connects them.
Operators lose money to problems software already sees.
Sensors are everywhere. But there's no AI team on site — so defects slip through, equipment goes down unplanned, and inspection stays manual. The data to prevent it is right there, unused.
- — Missed defects reach the customer
- — Unplanned downtime costs six figures an event
- — Skilled inspectors doing repetitive checks
Physical AI is starving for data that doesn't exist yet.
Labs ran out of text. They need the physical world — but worker video shows motion, not results, and simulation isn't real. There's no supply of data that proves what actually worked.
- — Worker video = what humans did, not the outcome
- — Simulation ≠ the real factory floor
- — No outcome labels, no reward signal
Manifold is the layer in the middle. One deployment feeds both.
Work in. Intelligence out. On repeat.
No robots to install, no data collection crew. We use the infrastructure a site already has and turn its daily operations into a self-improving system.
Plug into what's already there
A small gateway box pulls video from existing IP cameras and machine data from PLCs. Local buffering means nothing breaks when the site's internet does.
Deploy agents that do the work
Agents monitor equipment, schedule it, run inspections against the site's SOPs, and make the calls a person makes today — verified frame by frame.
Log every action with its outcome
Each task lands in an episode store as state, action, and result. Because the AI did the work and got graded by reality, the label comes for free. No manual labeling.
Retrain, then push back to the floor
Models learn from their own results and redeploy to the site. The site runs better, the data gets richer, and every new site compounds the loop.
Every shift writes the dataset.
While the plant runs, the episode store grows. Each row is an agent decision graded by what physically happened next — the reward signal physical AI has never had at scale.
- Ground truth for freeThe outcome is the label. No annotation pipeline, no human raters.
- Dense, not curatedRoutine checks and rare failures both land in the store — the tails are where the value is.
- Exportable on contractLabs pull scoped, de-identified episode sets through the data portal.
The factory pays for the AI. The labs pay for the data.
The same deployment produces two products at once. That's the whole unlock — the operator funds the work, and the work funds the data.
AI that runs your operation, live.
We integrate deep into the production flow and take on the work that causes the most expensive failures — then hand you a new revenue stream on top.
Data nobody else can make.
Every episode is an agent taking a real action in a real facility and getting graded by the real result. Outcome-attached, production-grounded, exclusive.
Watching people has a ceiling.
Others put cameras on workers and sell the footage. That teaches a model what a human did — never whether it worked. The data that matters only exists if AI does the work first and gets graded by reality.
- The facility is paid to be watched — a cost it tolerates.
- Motion without results. No reward signal to learn from.
- Nothing gets better on the floor.
- The facility depends on us — a system, not a cost.
- Real actions, graded by the real result. A true reward.
- The floor improves and the data compounds, together.
Built on the infrastructure sites already have.
Capture and deployment at the edge, an outcome-graded episode store in the cloud, and agents that reason against real standard operating procedures.
Runs at the site, survives the site
- · Python services on a small gateway box
- · Pulls IP-camera video + PLC machine data
- · Local buffering — nothing breaks when the internet does
- · Fine-tuned open vision models for fast perception
An episode store, not a data lake
- · Python on Google Cloud
- · Telemetry → Cloud Storage + Postgres
- · Every record is state · action · outcome
- · Retrained on our own episodes with PyTorch
Reasoning where it counts, speed where it doesn't
- · Claude Agent SDK for managed agents
- · Opus for complex reasoning + SOP verification
- · Sonnet for high-volume routine checks
- · Frames inspected against each site's SOPs
One floor view, one data portal
- · TypeScript · React · Next.js
- · Operator dashboards, alerts, work orders
- · Data portal where labs browse + pull datasets
- · Outcomes attached to every export
If physical AI follows the LLM curve, the data spend follows the compute curve.
Near-term: ~$45M/yr from a hundred sites and a handful of labs. The ceiling is far higher — we're building the supply of outcome-graded operational data before the demand fully arrives.
The details operators and labs ask first.
Something else on your mind? getmanifold@gmail.com
Do we need to install new cameras, sensors, or robots?+
No. We plug into the IP cameras, sensors, and PLC records your site already has. A small gateway box handles capture and keeps buffering locally when the internet drops. Deployment is software-first — no new hardware on the line.
What does the AI actually do on day one?+
It starts with the highest-ROI calls: monitoring equipment health, inspecting output against your SOPs, and flagging scheduling conflicts. Early on, agents shadow your team and every action routes through human review. As accuracy proves out on your numbers, they take the routine calls directly.
How is this different from putting cameras on workers?+
Worker video records what a person did — never whether it worked. Our agents do the work themselves, so every record carries the outcome: state, action, result. That's the difference between footage and a reward signal, and it's why the data commands a different price.
Who owns the data, and what do we earn from it?+
The episodes come from your operation, and you share in every sale — the operator takes a cut of the data revenue on top of the operational gains. Datasets are sold under contract, scoped per buyer, and de-identified so your proprietary process stays yours.
What does a pilot look like?+
Weeks, not quarters. We integrate with your existing systems, agents shadow the operation to baseline, then start taking real actions on defined workflows. The pilot converts to a paid per-site deployment when the value shows up in your numbers — that's the structure our current pilots run on.
We're an AI lab. What can we actually buy?+
Outcome-graded episode datasets from live industrial operations — and, for the right partners, access to the environments themselves: real facilities where policies can be evaluated against physical results. Contact us for the data portal and current coverage.
Put AI on the floor.
Turn the floor into data.
Two manufacturing pilots are live and converting to paid deployments. Whether you run a site or train physical AI, there's a seat in the loop.