What the product actually is
Inspection on a line is a judgment under cost asymmetry. A false reject stops a good unit: scrap, rework, throughput. An escaped defect leaves the plant: warranty, recall, a customer who does not come back. The model is a camera and a score. The product is which error you are willing to make, who can override it and whether the system still works on Tuesday night lighting with a new SKU.
That is why this work belongs next to Story Quality and AI Peer Review. Different domain. Same job: define correct behavior when the model is allowed to act.
The 1.5-month constraint
A 1.5-month MVP is not a flex. It is a scope weapon. You cannot take “detect all defects in manufacturing” to a plant in six weeks. You pick one inspection job that is frequent, expensive when missed and photographable. You put a human in the loop on purpose. You measure false reject and escape on that job, not on a slide titled “AI accuracy”.
Time-to-market dropped 50%. The useful reading of that number: we stopped treating the first release as a complete vision system. Completeness is how industrial AI dies in integration. A narrow, deployed inspector that a line supervisor can disagree with is worth more than a general model that is always almost ready.
Behavior on the line
I will not invent plant-level confusion matrices I cannot publish. The behavior I will stand behind:
- Fail visibly. If confidence is low, the unit is not “maybe OK”. It is held. Silence is an escape.
- Override is first-class. A supervisor override that is not logged is not an override. It is the model decaying in the dark. Overrides are the next training signal and the current safety valve.
- New SKU / new lighting is a product event, not an engineering surprise. Domain shift is the default. If the UI does not make “this is a new condition” cheap to flag, you will eat escaped defects for breakfast.
- Latency is a quality metric. An inspector that is right 200ms after the unit has moved is not an inspector.
Launch is part of the model
The launch work was unromantic and load-bearing: cloud deployment, positioning, acquisition, pricing. Five new customers. Compliance in parallel (SOC 2, ISO 27001 and GDPR) because a vision system that looks at a customer’s line is a data and access product whether the pitch deck wants that or not. You do not get to sell “AI for manufacturing” to a serious plant without being able to answer where the images go.
I also owned the website overhaul (+40% engagement, +25% lead generation). That is not the interesting artifact. I include it so the record is complete and so nobody thinks the only work was a model. At a 50-person industrial AI company, the TPM’s job includes the boring surfaces that make the real product findable.
On the round
Frinks.AI raised $5.4M Pre-Series A in 2025, led by Prime Venture Partners, with existing investors including Chiratae. I will say what is true: product success was a key factor. I will not say I raised it. Attribution inflation is how you fail the first diligence question in an interview.
What I would do next, given the same product
- Report false reject and escape as separate product dashboards, always. A blended “accuracy” number is how you hide the error the plant actually feels.
- Treat override clusters as a release blocker. If one SKU soaks overrides, you do not have a model problem in general. You have an unannounced product hole.
- Make the human-in-the-loop path faster than the informal WhatsApp workaround. If the official hold flow is slower than texting a photo to a supervisor, the official flow will lose.
- Price against escaped-defect cost, not against “AI”. Plants already have a number for what a miss costs. Use it.