AI got you a demo.
This is production.

AI-shipped prediction market. Production-grade, week by week.

Glimpse Markets launched fast on AI-generated code and inherited a slow, fragile trading frontend. I audited the full codebase in 48 hours — severity-ranked security holes, untyped boundaries, missing error states, and the performance bottlenecks that actually hurt traders. On go-ahead, I shipped the first week of frontend fixes in four days: killed the browser-side third-party fan-out, routed market data through their own backend, and cut first paint from sluggish to instant. It runs now as an ongoing weekly retainer.

"Sushil found a week of latency hiding in our frontend and shipped the fixes in days. First paint went from sluggish to instant." — Ruban, founder · Glimpse Markets

AI writes the program. Code gates the bad ones out.

Leverage is a strength-training app I built for my own program and use myself. The AI builder generates a full training block from goal, schedule, and equipment — then deterministic checks gate every program before it renders. Push and pull volume stay balanced, skill work sits first while you're fresh, and progressions advance only on held standards. An L-sit hold has to reach 20 to 30 seconds before the next step unlocks.

Hand-written vanilla JavaScript on PocketBase, no build step, installs as a PWA. Daily check-ins, streak tracking, weekly volume, and a guide layer that explains the program's reasoning in plain language. The AI generation is useful because I wrote the rules that govern it.

Strategy engines, versioned and benched like software.

Manja is my paper-trading desk for NSE equities. Seven forward books run on versioned engines: momentum on 1.31, a momentum-quality blend on 1.29, and unproven strategies sit unbenched until forward results earn them a book. Every fill carries modeled costs and gets marked daily.

Breadth, sectors, and sentiment read in one view, on dark and light grounds. No real money rides on it yet, by design: strategies graduate on forward evidence, not backtests. It is the same evaluation discipline I put around AI features before they ship anywhere. Build it right, gate it properly, then run it in production.

Three more builds in motion.

Open. Could be yours.

Available for one engagement starting June 2026. Your AI-built MVP, an agent build that needs guardrails, or a codebase that needs to graduate from demo to production — let's scope it.