We closed initial angel funding in September 2025, and this feels like the right moment to write down where we are and where we're headed — not as a press release, but as a straightforward account for the quality engineers and operations teams we've been working with since we started.
Procunit has been operating for two years. We started with a question that Amanda had been sitting with from her previous job: why do rule-based AOI systems fall apart when you speed up the line or change the part? The answer turned out to be simpler than the industry wanted to admit — rules written by engineers for a production environment that no longer exists once conditions shift. The model-based approach we built from the start handles that adaptation by design, not by exception.
Where we are today
We've run inspection pilots on stamped metal components, machined housings, injection-molded plastics, and wiring harness connectors. Every pilot started the same way: one camera, a 50-image training set, an overnight training run on the customer's industrial PC, and a two-week measured comparison against their existing inspection process.
Not every pilot converted to a paid deployment. Two didn't — one because the customer's part volume was too low to justify the per-line cost (we were honest about that), one because the defect classes on their die-cast aluminum line required a second camera angle we didn't have a clean mount solution for at the time. Both are fair critiques and we've taken them seriously.
The pilots that did convert validated the core hypothesis: a quality engineer with no ML background can go from labeled images to a live inspection signal in one working day. The team that surprised us most was at a stamping operation with four quality engineers, none of whom had touched any kind of vision software before. They were retraining models themselves within three weeks of go-live. That's the version of the product we wanted to build.
What the capital is for
Angel capital at this stage is deliberately specific. We're not trying to expand to new verticals or build out a sales team. The money goes to three things.
First, multi-camera line support. Right now, Procunit handles up to four synchronized camera channels on a single inference node, but the orchestration between cameras — correlating detections across views, handling part rotation between stations — requires more engineering work than we've been able to give it. Several customers have asked for this and we've had to defer. That changes over the next several months.
Second, the Procunit Labeler. The labeling interface works, but it's basic. Quality engineers have told us they want keyboard shortcuts that experienced annotators expect, better handling of difficult lighting conditions during annotation, and a way to review borderline cases as a batch before they go into the training set. These are all reasonable requests that have been sitting in the backlog. We're going to clear them.
Third, deployment support infrastructure. Running one pilot at a time is manageable. Running several simultaneously, with different hardware configurations at different facilities, requires better remote diagnostics and update delivery than we have now. We're building that so we can scale pilots without adding headcount proportionally.
What we're not building
We're not building a cloud-hosted inference offering. Every manufacturer we've talked to has some combination of data residency concerns, network reliability concerns on the production floor, or latency constraints that make cloud inference impractical for real-time reject decisions. On-premises inference is not a temporary constraint we're working around — it's the right architecture for this problem. We're not changing that.
We're also not expanding to new industries right now. We know discrete manufacturing. We know the specific failure modes in stampings, castings, and assembled components. Adding a new vertical — medical device inspection, food and beverage, semiconductor wafer handling — would require us to rebuild our understanding of defect physics for that material class. We're not ready to do that credibly at this stage, and we're not going to fake it.
There are already products that try to cover every inspection use case and end up doing none of them well. That's the legacy AOI playbook. Rule-based, one-size-fits-all, and expensive to reconfigure. We built Procunit specifically to be the alternative for the discrete manufacturing use case. Staying focused on that is how we stay useful.
The roadmap for the next 12 months
Multi-camera orchestration ships first — it's the most requested capability and the one with the clearest path to implementation. After that, labeler improvements, then the deployment infrastructure work. If multi-camera support proves out as well as we expect, we'll be in a position to address the cylinder bore and internal surface inspection requests we've been getting, which require radially-mounted camera arrays and different illumination setups than our current line-scan architecture handles.
We're also doing more careful work on escape rate measurement tooling within the dashboard. Several quality engineers have told us they can see their detection data but struggle to produce the audit-closed escape rate numbers their quality managers want. We wrote about that measurement problem on this blog recently. Building the audit reconciliation workflow into the product directly is on the list.
A word on where we're not
We're a small team. We've been cautious about the pace we set during the pilot phase because credibility in this industry is harder to earn back than it is to spend down. When something doesn't work — a false positive rate that's too high on a particular surface texture, a model that drifts after a tooling change — we'd rather diagnose it honestly than paper over it with optimistic presentation. The manufacturers we want to work with are the ones who share that standard.
If you've been waiting on a pilot because you weren't sure the product was real or the team was serious: the capital helps, but the more meaningful signal is the deployment history. We're not asking anyone to take a bet on promise. We're asking for a two-week pilot on one line. If it doesn't perform, you haven't committed to anything.
The next post will return to the technical content we've been writing — camera placement for cylinder bore inspection is queued up. But we thought this moment warranted a direct update. Thanks to everyone who's run a pilot, given us honest feedback, and pushed back when something wasn't working.