Case studies.
Four deep dives into how I think about building with AI — from build-vs-buy architecture calls and privacy-first data pipelines to a behavioral-AI teardown you can play. Each one is a decision, not just a demo: the constraints, the trade-offs, and the reasoning behind the choice.
Should this site's AI run a self-hosted LLM?
A structured build-vs-buy evaluation of the “Ask Gazzali.ai” terminal: model selection, quantization, a RAG layer vs. fine-tuning, an eval harness, and the cost crossover that decides it — ending in a disciplined “buy now, build past the crossover.”
Read the teardown →MyHPTracker — private by architecture
An OCR-free, multimodal-LLM pipeline that turns messy medical documents into a queryable health profile — with privacy designed into the architecture and provider-agnostic AI that cut inference cost 6.5–17×.
Read the case study →TenderAI — tender deviation analysis
A full-stack Next.js + FastAPI + Gemini platform that flags deviations in tender contracts and drafts clause-anchored responses to queries — with a human always kept in the loop.
Read the case study →Rock, Paper, Scissors & Behavioral Patterns
A camera-based Rock·Paper·Scissors that learns a quirk you don't even know you have — a teardown of behavioral-pattern personalization (and why it's a gold mine for B2C), with the game playable right inside the page.
Read it & play →See the AI run.
The assistant these studies are about — live on the résumé.