The problem
Picture a first-time investor in a Tier-2 Indian city. A company she holds announces a block deal, a bonus issue, or quarterly results. The information is public within seconds — but it arrives as a dense English regulatory filing, written for professionals. By the time she understands what happened, the professionals have long since acted. India has tens of millions of such investors, and almost all market information reaches them late, in the wrong language, in the wrong form.
The solution
Live feeds from both Indian exchanges flow into a pipeline that classifies each event — corporate actions, block and bulk deals, insider trades, results, IPO updates, mutual fund changes — using a deterministic matcher trained on 4,887 listed-company name variants, so a headline is never attributed to the wrong stock. An AI layer built on Anthropic's Claude writes a short, plain-language insight; a translation and text-to-speech layer speaks it aloud in Indian languages, targeting under half a second from feed to audio on the fastest path.
Because regulation matters, the system informs and never advises: a classifier checks every generated sentence and fails closed — if it even resembles investment advice, it does not go out. The system runs today — feed, classifier, AI insight engine, mutual fund and IPO pipelines, mobile-style app, paywall and all — with more than six hundred logged, verified changes in its delivery ledger across roughly eight weeks of build.
Behind it sits our model-engineering practice: teacher–student model distillation — large voice models teach small per-voice students that speak at 84 milliseconds per sentence on a modest self-hosted GPU — plus LoRA fine-tuning and per-language pronunciation dictionaries so speech models say financial vocabulary correctly across ten Indian languages, with self-hosted small models benchmarked head-to-head against hosted frontier models before anything ships.
