Shwetal Mehta

Three outcomes

What the work produced, before how it was built.

Case studies

What was broken, what I built, the numbers as measured, and what didn't happen.

• Four teams to one seller An AI-assisted pricing and SOW engine, built solo. Pricing days to minutes; cost per SOW ~13× lower; zero legal rework.
• Aria: designing against engagement A design study of an AI wealth manager: the product thinking, and what an audit a year later found the decision log had claimed but never built.
• NovaHub and NovaQuant: the system that paused itself A market-data and ML platform, an autonomous trading system, and the controls that stopped it when a review said stop. Directed against written specs; repositories stay private.
• sma-cockpit: a sixty-second options research desk Scans an options universe every minute, ranks candidates on expected value per unit of tail risk, and grades its own output nightly. Backtester and live grader share one code path.
• compass: finding the edge between P and Q Our conditional terminal distribution against the one the chain is quoting, priced in dollars per contract — and the census that decided what a rebuild was allowed to inherit.
• Onshore to automation-first A delivery model rebuilt twice in seven years. States plainly which figures are measured and which are recalled.

Applications you can open

Built to be used, not watched. No sign-up, no demo request.

Demonstrate to Win

Teaches sales engineers to demo AI products, where confidence scores invite doubt and "it learns over time" sounds like "it doesn't work yet". Eleven modules, 31 activities, two certification tracks.

Source · React 19 · no backend

AI Work Simulator

Who does the world's work by 2036? Six regions, six task classes, eight assumption dials. Move a dial and the verdict rewrites itself. Every scenario is a URL you can send.

Seeded with ILO, IMF, IFR, METR and Census data

Security Posture Simulator

Toggle fifteen controls and watch risk move across nine threats. The same computed state is then argued four ways, for a CISO, a CFO, an architect and a SOC manager.

Source · tested scoring engine, 36 tests

PlateLens

A proof of concept built in one sitting: can a model look at a meal photograph and reason about physical size? It hunts for an object of known size in frame to anchor scale. Not finished, not production, no accuracy claims — the point is that this is AI doing something that isn't a chatbot.

Recorded run, ten photographs, mistakes included

The Tier Ledger

AI capability on a human-anchored 0–5 scale across eight domains, three readings each: what the lab has, what you can buy, what the economy actually runs. The gap between those three is the interesting part.

Eight half-years scored, three projected

Open source

The method, extracted from the systems below and stripped of their domain.

• ai-system-governance Doctrine, retirement checklist and a runnable promotion board for keeping an LLM system honest. 47 tests.
• spec-driven-ai-development Specify before you prompt, log what you asked, number the documents. Templates and a worked example.
• astrasim Monte Carlo simulator scaffold. Public for its prompt-ID log and decision records, not its engine.

Private systems

Four private systems, three running and one deliberately retired into its successor. They stay closed because they carry the trading logic that was the point of building them. The engineering discipline does not have to stay private, so every figure here is counted from the repository. Each one links to its case study — what it was for, how it works, and what it cost to learn. The repositories themselves stay closed.

Three rules

Earned from specific failures in the systems above, not adopted from a framework.

What's next

Kept off this page on purpose. A portfolio that lists intentions beside results teaches a reader to discount both.