Vision for chemical procedures. It watches the apparatus, tracks it in real time, announces the moments that need a hand, and keeps the record.
What is LabVision?
LabVision is a vision system for chemical procedures.
Put a camera on the bench and it watches the glassware, follows the liquid, and says out loud the moment something needs a hand — while the procedure is still running.
The first procedure is built: draining a separatory funnel, where it tells you when to close the valve. Titration is next.
Measured, not estimated
Our own measurements of our own system on our own footage, made in-house in September 2026 on the funnel procedure. Every number comes from a run folder that ships with the command that produced it. What they mean, and what they do not.
How we do it
A camera does not wait. Everything in the system is built around that one fact, and every timing number on this page depends on it being true. Frames are never queued; a decision never waits past its budget; and where the system cannot see, it says so instead of guessing.
Frames arrive at the camera's own rate. If the system is not ready for one, that frame is dropped and counted. Falling behind costs frames, never time, so an announcement is never late because of a backlog.
Our trained perception model finds the glassware and the liquid in every frame while the procedure runs. Each result is stamped with its age. Nothing is precomputed; the video is understood as it happens.
If the outline it is measuring against is older than 1.2 seconds, the system refuses the measurement and goes quiet. A decision never waits longer than 600 ms; the median is 118 ms. In 20,092 refusals, not one carried a value.
Where the milliseconds go
| Arm | Decided frames | Median | 90th pct | 99th pct | Worst frame | Over 600 ms |
|---|---|---|---|---|---|---|
| Perception model in the loop, quiet machine | 4,634 | 118.4 ms | 193.1 ms | 237.3 ms | 391.3 ms | 0 |
| Perception model in the loop, machine half busy | 3,964 | 30.1 ms | 277.6 ms | 434.4 ms | 572.1 ms | 0 |
| Perception replayed from disk, quiet machine | 5,855 | 11.4 ms | 140.7 ms | 192.1 ms | 270.4 ms | 0 |
Cumulative share of decided frames answered within a given time, six development videos replayed at real speed on one laptop. The dashed line is the 600 ms deadline. On a half-busy machine the tail stretches to 572 ms and the system announces less than half as often: load does not make it late, it makes it quiet — the safe failure for a system whose job is to say "close the valve now".
The three parts of the median frame; they sum to the latency by construction. The perception model contributes nothing to this number: it runs on its own lane and its cost is charged to the age of the evidence, not to the answer.
Procedures
The camera watches the boundary between the two liquid layers approach the stopcock and tells the chemist when to close the valve: a warning first, then a confirmation once the lower layer has left the cone.
Request the demoThe camera watches the flask for the first colour change that stays, and announces the endpoint so the burette is stopped on the right drop.
The camera watches the solvent front climb the plate and says when it is time to take the plate out of the chamber.
The camera watches the solvent level above the silica, warns before the column runs dry, and says when a fraction tube is full.
The camera sees the ring of condensation in the condenser, starts the reflux timer when it appears, and warns if the ring climbs too high.
The camera checks that the vacuum was broken before the pump was switched off, and notes when the filter cake has stopped dripping.
The camera watches the receiving flask fill and stop, and flags a bump into the trap.
The camera confirms that liquid actually moved between vessels, not just that the motion was made, and logs the transfer.
The camera watches the pot level, the stirring vortex and the plate so nothing is heated dry or left unattended.
Every announced event is timestamped and written to the record, so the account of a procedure is made by the camera, not typed in afterwards.
Why we do it
Every laboratory procedure ends the same way: somebody writes down what they did. In a regulated lab a second person stands beside them so the record can be trusted. That record is what an inspector reads, and writing it is time no scientist was hired for.
So we put a camera on the bench and let it do the writing. It watches the procedure, announces the moments that need a hand, and keeps a time-stamped record that can flow into the lab's notebook or quality system. The same record lets a new operator ask "is this being done right?", and later gives a robot arm the eyes it needs to do the procedure itself.
The figures below are other people's published research, not ours. None of them says anything about how well our system works; for that, see the evidence.
FDA warning letters to drug and biologics makers in fiscal 2025 — 59 % more than the year before. Batch-record review was cited in 47 of them.
of quality-control labs still run on disconnected systems with limited automation. Half of those that modernised report fewer errors and deviations.
for a second person to check one medication — 107 hours per thousand. Where the check was required it made no measurable difference, because almost none were truly independent.
of the good-practice cost of a phase III trial goes to one activity: a person checking that what was written down matches what happened.
Tudur Smith et al., PLoS ONE, 2012
lower cost for a digitally enabled chemical quality-control lab, with more than a 65 % reduction in deviations.
a week saved per researcher after adopting an electronic lab notebook, and a third less time spent on reporting. Seven-user study, published by the vendor.
a year: computer vision in healthcare, from USD 3.93 billion in 2024 to USD 14.39 billion in 2030 — the technology this product is built on.
a year: smart glasses, from USD 1.22 billion in 2025 to USD 4.13 billion in 2030 — the hardware for a camera that travels with the operator.
Nine neighbouring markets
Laboratories already pay for software that stores what they did and for automation that does it for them, and both budgets are growing. The two fastest-growing neighbours — computer vision in healthcare and industrial smart glasses — are the technology and the hardware this product is built on. What none of those products does yet is perceive the procedure itself. That is the piece we build.
Each market at its base year and its published forecast, with the annual growth rate the publisher reports. The forecasts are theirs, not ours. Sources: MarketsandMarkets (machine vision; lab automation, May 2026; laboratory software, Dec 2025; computer vision in healthcare, May 2025; pharmaceutical manufacturing software; pharmaceutical QMS, Nov 2025; smart glasses, Sep 2024; ELN, Jun 2025) and The Business Research Company (laboratory robotics, 2026).
Evidence
To us a benchmark means the best published models and ours on the same footage, held to the same requirements, scored against ground truth marked by a person. That comparative benchmark is being built; it needs ground truth that mostly does not exist yet. What is here is narrower, and we say so plainly: the requirements were written down before the code, and every number sits in a run folder with the command that made it.
Six requirements, written before the code
Separatory-funnel procedure, six development videos, replayed at real speed.
| What we required | Held? | What we measured |
|---|---|---|
| Every frame is either processed or counted as dropped. None goes missing. | YES | scheduled = processed + dropped, exactly, in all 14 runs |
| An announcement follows the frame that justifies it within 200 ms in the typical case and 300 ms in the slow tail. | YES · 6 of 6 videos | worst median 178.9 ms · worst 95th percentile 270.5 ms |
| No decision is made before the evidence for it has arrived. | YES | 0 violations in 22,513 decisions |
| Memory stays flat over a long run. | YES | +22.3 MB across a 630 s replay |
| When the system refuses to answer, it gives no value at all. | YES | 0 of 20,092 refusals carried a value |
| Under load it drops frames, not deadlines. | YES | every drop counted · 0 announcements past the limit |
Why we trained our own model
The question at the bench is not accuracy on a saved picture. It is whether the system says anything while the clock is running and it has to pay for its own computation. We put a large, published, general-purpose vision model in the live loop in place of ours.
| What sees the frames | Announcements | Decisions |
|---|---|---|
| A large published vision model, live | 0 | 1,027 |
| Our trained perception model, live | 3,159 | 11,803 |
Each of the large model's results took about 2.1 s to compute. By the time it landed it was older than the 1.2 s the tracker allows, so every result was thrown away and the system stayed silent for the whole run. Same six videos, true real time, nothing computed ahead of time. The small model is not an optimisation; without it there is no system.
On photographs it had never seen
107 unseen photographs of separatory funnels with hand-marked outlines. Median distance from the drawn outline to the true glass edge, and how much of the liquid–liquid boundary was found.
| Model | Outline error | Boundary found | Per frame |
|---|---|---|---|
| Our trained perception model | 5.0 px | 92 % | 45.7 ms |
| A published compact model, trained the same way | 5.4 px | 86 % | — |
| A large published vision model, untrained on funnels | 11.2 px | 48 % | 341 ms |
Three training seeds averaged for the first two rows. Only the boundary-coverage gap is statistically clean; the outline errors of the first two rows overlap and should not be quoted as a difference. The large model has never been shown a separatory-funnel label, which is the point: a generalist is not enough, and a specialist need not be big.
What does not work yet, stated plainly
The moment the system calls "close the valve" is not yet stable from run to run. On one video it fired at 9.08 s where a person had marked 11.72 s, and a second run of the same configuration the same afternoon fired at 6.96 s. On a newer clip, seven equally trained versions of the model called the valve anywhere from 8.1 s to 19.4 s, and two never called it at all — and nothing can adjudicate between them, because that clip has no ground truth. On the best video the system reached a decision on only 26 % of frames under the current rules. Two- and three-layer separations work; a fourth layer is not reported. The development videos are 24 to 30 frames per second, not the 60 the requirement names. Nothing on this page is a claim until it sits in an experiment folder with the command that made it — and today that is 171 folders and 880 automated tests.
Questions we are asked
Every number below was measured in-house in September 2026 on the funnel procedure.
Because the record is the bottleneck. Writing down what happened, and having a second person verify it, costs a regulated lab real time, and getting it wrong costs more — FDA warning letters to drug makers rose 59 % in fiscal 2025, with batch-record review among the most-cited failures. A camera that perceives the procedure can make that record itself. The numbers and their sources are in why we do it.
A camera on the bench, on a stand, at an angle that gives it a clear view of the apparatus. Nothing is attached to the glassware and nothing about the procedure changes. The camera feeds a computer with a graphics card; our measurements were made on a laptop with an RTX 5070 Ti.
The live footage with the tracking drawn on it, and the announcements in words. On the funnel procedure that means the outline of the funnel, the fill line, the boundary between the layers and the valve line drawn on the picture, and two messages: a warning ("close the valve in about 0.6 s") and then a confirmation once the lower layer has left the cone. When the system cannot see well enough, it draws nothing and says nothing, instead of showing a line it is not sure of.
Two things: the timestamp of every announced event, and the annotated recording of the run. The plan is to write both to storage you already own and choose, such as your lab's own drive or electronic notebook, so the record never has to leave your site. That storage link is planned, not shipped; today the events and the recording are written to files on the computer running the system.
One is built: the separatory-funnel drain, which is what the demo shows. Planned next, in order: titration, thin-layer chromatography, column chromatography, reflux, vacuum filtration, rotary evaporation, liquid transfers and pipetting, distillation and hot-plate heating, and a compliance log across all of them. Each is described in one sentence in the procedures.
Apply on the request page with your name, organisation and email. We read every request by hand. When it is approved you get one email with a private link that works for a set number of views. The demo is a recorded run of the funnel procedure: the original footage, what the system sees, and what it concludes, side by side, with the facts and caveats written under it.
Not to start: the funnel procedure works as shown without any footage from you. But every bench is different in its light, camera position and glassware, and the system can be adapted to a lab's own footage so that it learns that bench. We treat that adaptation as part of the product rather than a workaround. We do not yet have numbers on how much it helps, so we make no claim about that here.
Yes, in principle, because the system sees and decides in real time. Measured in-house on the funnel procedure: our trained perception model takes 45.7 ms per frame in the typical case and 64.1 ms in the slow tail on a laptop RTX 5070 Ti; the tracking code adds 19 to 40 ms per frame; and the announcement decision has never exceeded its 600 ms budget in 25,476 decided frames, with a median of 118 ms.
What it supplies is the task-level signal, not the motor signal: where the boundary is, when to close the valve, a warning and then a confirmation, about 20 times a second. A robot arm's own motor loop runs hundreds of times a second and this system does not feed it. Connecting it to an arm's controller is future work, not a shipped feature.
Beyond the procedures themselves, two things the same camera makes possible. First, a compliance record: because every event is announced with a timestamp, the account of a run is written by the camera as it happens, not by a person afterwards. Second, a vision layer for robots doing chemistry. Both grow with each procedure added to the list; neither is a separate product.
A recorded run of the funnel procedure — the footage as filmed, what the system sees, and what it concludes — behind a private link that works for a set number of views. We read every request by hand.
Occasional notes when a procedure reaches the bench or a measurement changes. No cadence promised, no list sharing.
Contact
We reply to everything.