work
AI, backend, and the systems underneath.
- The dashboard holds 4 layers of climate and crime data across 650+ districts(India and Sri Lanka), but you had to know where to look. With the AI assistant chatbot (built from scratch), you can now ask a question in plain words and get reliable answers back, any district, any year.
- You can speak instead of type, in any of seven regional languages, and it works out which one you used. Built a benchmark harness that checks speaking gets you the same answer as typing, holding transcription error near 2%.
- Video ads cost more than these brands could spend. Built a prompt-to-video pipeline — Gemini for the storyline, Veo 3.0 for the shots — that turned out 60+ finished ads at 2–3 a day.
- Volume only helps if the work doesn't come back. Brand voice went into the prompts themselves, so ads shipped with almost no revision rounds and pulled 3K–17K organic views a post.
- An avatar that looks different every run can't carry a product. The multi-modal AI avatar system I built with the CEO holds one identity through a prompt refinement pipeline; inconsistencies down 30%, iteration time down 25%.
- Fidelity was the blocker. 700+ prompts tuned across Stable Diffusion, Midjourney and Gemini lifted visual quality 40% over baseline, and settled which models the system runs on.