ChemLabs is a vision system for chemical procedures. Put a camera on the bench and it watches the apparatus, tracks it in real time, announces the moments that need a hand, and keeps the record. The first procedure is built and shown today: draining a separatory funnel, where it tells you when to close the valve. Titration is next.
All four are our own measurements of the funnel procedure, made in-house in September 2026. See the benchmark section for what they mean and what they do not.
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 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 of the funnel it is measuring against is older than 1.2 seconds, the system refuses the measurement and goes quiet rather than guessing. And a decision never waits longer than 600 ms; the measured median is 118 ms. In 20,092 such refusals, not one carried a value. What you hear is either current or absent.
To us, a benchmark means this: the best published models and ours, on the same footage, held to the same requirements, scored against ground truth marked by a person, and kept running as models change. That comparative benchmark is being built. It needs ground truth that mostly does not exist yet, and we are collecting it.
What is on this page today is narrower, and we want to say so plainly: these are our own measurements of our own system on our own footage. The requirements were written down before the code, and every number comes from a run folder that ships with the command that produced it. They show that the system is fast and honest about what it sees. They do not yet show how it compares with anyone else's.
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 typical (median) 178.9 ms · worst slow tail (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 |
The question that matters at the bench is not accuracy on a saved picture. It is whether the system says anything at all while the clock is running and it has to pay for its own computation. We tested that by putting a large, published, general-purpose vision model in the live loop in place of ours.
| What sees the frames | Announcements | Decisions | What happened |
|---|---|---|---|
| A large published vision model, running live | 0 | 1,027 | Each result 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. The system stayed silent for the whole run. |
| Our trained perception model, running live | 3,159 | — | The same six videos, at real speed, with nothing computed ahead of time. |
Both runs used the same videos at true real time, with no speed-up and every input paid for as it arrived. The small model is not an optimisation; without it there is no system.
The moment the system calls "stop draining" 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 the best video the system reached a decision on only 26 % of frames under the current rules. The demo clip has no ground truth at all: the timings shown in it are the system's own readings, not scores. And the development videos are 24 to 30 frames per second, not the 60 or more the requirement names. Nothing on this page is a claim until it sits in an experiment folder with the command that made it.
One procedure is built. The rest are planned, in the order below: the ones a camera can be taught first come first. For the planned ones we say what the camera would watch and announce, and nothing about how well, because none of them has been built or measured.
Every laboratory procedure ends the same way: somebody writes down what they did. In a regulated lab a second person often 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.
The cost of getting the record wrong is rising. The US FDA sent 303 warning letters to drug and biologics makers in fiscal 2025, 59 % more than the year before, and one of the four most-cited failures was the review of batch records and the investigation of discrepancies. Data integrity has been the most common theme in those letters for years.
So we put a camera on the bench, or on the operator, 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.
Somebody already does this job by hand, and it is measurable. A second nurse checking one medication takes six and a half minutes, and in the study that watched them, the check the rules required caught nothing — because a person standing beside a colleague is almost never checking independently. In clinical trials, half the cost of good-practice work is one person verifying that the record matches what happened. An instrument that watches does not get tired, and it does not know what answer it is supposed to see.
We are not inventing a market. 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.
303
FDA warning letters to drug and biologics makers in fiscal 2025 — 59 % more than the year before. The batch-record and discrepancy-investigation rule was cited in 47 of them.
ELIQUENT Life Sciences, “Trends In FDA FY 2025 Warning Letters”, Pharmaceutical Online, March 2026
40 %
of quality-control labs still run on disconnected systems with limited automation. Half of those that modernised report fewer errors and deviations. Survey of 103 biopharma executives.
25–45 %
lower cost for a digitally enabled chemical quality-control lab, with more than a 65 % reduction in deviations.
9 h
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.
SciNote, “Return on investment when implementing an ELN in your lab”
6.4 min
is what it takes a second person to check one medication — 107 hours per thousand. Where the check was required, it made no measurable difference to the error rate, because almost none of those checks were truly independent.
50 %
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
$100 bn
a year is what McKinsey puts on the potential value of AI in pharmaceuticals and medical products — 4.2–6.1 % of industry revenues, among the highest shares of any sector they measured. This is AI in general, not computer vision.
24.3 %
a year: computer vision in healthcare, from USD 3.93 billion in 2024 to USD 14.39 billion in 2030.
29.4 %
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.
The figures above are other people's published research, not ours. Nothing here claims anything about how well our own system works; for that, see the benchmark section, which says exactly what we measured and what we did not.
Short answers. Every number below was measured in-house in September 2026 on the funnel procedure, and nothing is claimed that was not measured.
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 a camera on the bench.
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 section.
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 model sees, and the system's output, 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 in September 2026 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 on our bench clip; and the announcement decision has never exceeded its 600 ms budget in 25,476 decided frames, with a median of 118 ms. On the demo clip the warning came 0.75 s before the lower layer left the funnel and the confirmation 0.71 s after the warning.
By supplying the task-level signal, not the motor signal. A robot arm's own motor loop runs hundreds to a thousand times a second, and this system does not feed that loop. What it supplies is what a person at the bench would supply: where the boundary is, when to close the valve, a warning and then a confirmation. It does so about 20 times a second, with a decision inside 600 ms, which is the pace at which an operator or a task planner acts. 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, as described above. Both grow with each procedure added to the list; neither is a separate product.
Occasional notes when a procedure reaches the bench or a measurement changes. No cadence promised, no list sharing.
Questions about the method, the numbers, or running this on your own footage. We reply to everything.