Work & impact

Fourteen systems, and what each one cost.

Healthcare AI, enterprise ERP, camera-driven infrastructure, multi-tenant SaaS. Each one is written up the same way: the situation, the calls I made, the trade-offs I accepted, and the outcome. Several include the load-bearing decision as code.

Built at

Industries

Showing 6 of 14

Healthcare AI

World's Health Platform

In development. The architecture and the governed pipeline are designed, and the platform is being built slice by slice, each one leaving the full test suite green. Every clinical record in it so far is synthetic.

A cardiovascular care platform where AI drafts and physicians decide. Lab results run through a deterministic interpretation engine, an LLM writes the physician brief from a pseudonymised packet, and nothing reaches a patient until a physician signs it. In development.

Bounded contexts

Identifiers sent to the model

Healthcare AI

AI Orthopaedics Tool

Pre-operative planning time fell by roughly 40% on the first ward it shipped to, and every case on that ward goes through the tool now. The client measured that 40%, not me. 5,000+ clinical cases have been planned in it.

Hip-replacement planning for the direct anterior approach. Detection places the landmarks, the overlay goes on automatically or by hand, and the surgeon adjusts size, orientation and femur rotation before signing anything. It computes delta leg length and offset. Planning that took hours takes minutes.

5,000+

Clinical cases

AI & Computer Vision

Smart Parking System

100,000+ vehicles since launch. Zero double-bookings recorded.

A camera-driven parking system handling 100,000+ vehicles across a 6-story garage, 75 cameras. Zero double-bookings since launch. That came from a database constraint replacing the concurrency check that used to sit in application code.

Booking Errors

100,000+

Vehicles tracked

AI Tooling

Chitra AI Image Generation

The root cause was one line of configuration, not a line of code — the live backend was running the stub provider. Diagnosing it took one request to the options endpoint, which reports the active provider and model precisely so this question has an answer. The fix was smoke-tested with an inline override before the environment file was changed, so a bad token would have left the deployment exactly as it was.

Text-to-image generation with a private, paginated history. React front end, Django API, and a provider credential that never leaves the server. Shipped with a swappable provider interface — which is also how it broke in production.

Env line that broke production

Provider keys in the browser

AI Tooling

LLM Chatbot with Streamed Replies

Streaming carries structured events rather than raw text, so the client can tell a finished answer from an abandoned one and report usage honestly. The decisions and their consequences are written down as ADRs in the repository, including the ones that were forced by the API rather than chosen.

A Gemini-backed chatbot with token-by-token streaming, Markdown and syntax-highlighted code, and conversations that persist per user. The interesting decisions are in the transport and in what the model bills you for but never shows you.

~3s

Time to first token

ADRs recorded