Three months in. Time to write down what actually happened, before the story gets cleaned up in retrospect.
Where we started
The founding moment was the end of the 50-conversation validation process I wrote about in October. Eight letters of intent across different organization types. A clearly defined problem with a clearly defined solution requirement: real-time visual authentication that works in logistics conditions with a false positive rate low enough to be operationally safe.
We knew the problem. We knew what good looked like technically. We had the CV background to build it. What we didn't have was a team, a legal entity, a data strategy, or a patent filing.
Ninety days to get all of that started simultaneously.
The team problem
Computer vision at production scale requires people who have done it before. Not researchers who have trained models in notebooks. Engineers who have deployed CV systems into environments where the input is messy, the latency matters, and the model has to keep working when the lighting changes.
This is a small population. Finding them takes time. We moved fast by being specific about what we needed rather than posting a generic ML engineer role. We were looking for people who had dealt with distribution shift in production, who understood the gap between benchmark accuracy and production accuracy, and who could work in a very early-stage environment where the infrastructure doesn't exist yet.
We hired two people in the first 90 days. Both fit that profile. Both had production CV experience. The bar was high and we didn't compromise on it.
The patent question
Anti-counterfeiting in logistics is a crowded space in terms of existing approaches. Physical tags, holograms, RFID, barcode systems. What we're building is different: model-based visual authentication without physical modification to the product. We think it's novel in the specific combination of detection architecture, deployment context, and authentication pipeline.
The decision to file early was straightforward. If we're right that the approach is novel, the patent creates a defensible position. If we're wrong, we lose the filing cost. The filing cost is small relative to the downside of building something patentable and not filing.
The technical specification for the filing is mostly complete. The claims are being refined with the IP attorney. We'll file in Q1.
The data problem
Training a model to authenticate products requires images of authentic products and images of counterfeits. Authentic product images: available. Counterfeits: not widely available for obvious reasons. This is the data problem everyone in anti-counterfeiting faces.
Our approach is a combination of synthetic data augmentation and data collection agreements with brand partners. The brands have inspection records with images. Those records are the most valuable dataset in the space and they're locked inside enterprise systems. The LOI relationships are partly about product validation and partly about data access.
The model performance on day one will not be the model performance at month twelve. The data compounds as the system runs in production.
What the first 90 days actually felt like
Fast and incomplete. Every decision made at speed produces some decisions you'd make differently with more time. The legal structure, the equity split, the founding team agreements: all of these were done quickly and I've already found things I'd have worded differently. None of them are wrong. All of them are locked in.
The pilot conversations with the LOI companies are progressing. The first pilot is being scoped for Q2. The technical architecture is defined. The patent filing is weeks away.
Three months in, this looks like a company. Three months ago it was a hypothesis.
With gusto, Fatih.