Turning founder chaos into shippable AI products.

Problem

ChatlyAI had decent retention for an AI chat product but inconsistent growth — the kind of user graph that goes up and then flattens. Up and flattens. The product worked. Users who stuck around loved it. The problem was getting people to stick around.

Constraint

A narrow high-growth window and no time to buy the answer. SEO content took 60 days to show results, which was too slow for the moment we were in.

Decision

I went after the activation moment rather than the top of the funnel. Users who joined an active thread within their first 10 minutes had dramatically higher 30-day retention, so we made threads persist — conversations that grew and built context over time, instead of every session being ephemeral. We shipped it without announcing it, because the bet was that users would discover it and not want to leave.

Outcome

200K to 1.2M+ monthly actives in under a year. D30 retention improved 2.3x over eight months, session length rose 40%, and the work contributed to $6.5M+ in revenue.

Problem

An Amazon-partnered fiber optics supplier in Aurora, Illinois was building about 240 cables a month. Before a cable can be terminated, its 250-micron fibers have to be in the right order and precisely aligned, and sorting the fibers for one cable took about three minutes.

Constraint

Each fiber is a quarter of a millimetre across and alignment had to hold to 5 micrometres — on a production line run by operators rather than data scientists, where every second of latency is paid on every cable.

Decision

I treated the model as the small part of the problem — maybe 30% of the work. The rest was everything around it: image capture that behaves the same on every shift, inference that fits the machine’s timing, and the edge cases a clean dataset never contains. A 96% model an operator cannot run is worth less than a 90% model they can.

Outcome

Fiber sorting fell from about three minutes to eight seconds per cable at 96% accuracy, with alignment held to 5 micrometres. On the organizer line, AWS cable output went from one an hour to six; across the wider factory automation, monthly capacity went from 240 cables to 9,600 — 40x. The machine is covered by viaPhoton’s US Patent 12,001,073 B2.

Problem

Smart carts need to understand what shoppers are doing in real time — picking up items, placing them back, weighing produce — and they have to do it in a live supermarket, not a lab.

Constraint

The hardest part was never accuracy. It was latency, on constrained hardware. Any delay between a shopper picking up an item and seeing it appear on the display destroys the experience, and the compute was a Raspberry Pi module inside the cart.

Decision

I chose MoViNet — a mobile video network built specifically for temporal understanding on edge devices — over a heavier, more accurate backbone, and then bought the accuracy back through data rather than parameters: 8,000+ action clips collected in-store rather than pulled from a public dataset.

Outcome

95% real-time accuracy across six action classes at sub-80ms end-to-end inference. Deployed and operational at Al-Meera Supermarkets in Qatar, with real shoppers.