We build conversational AI systems that people reach on WhatsApp — and we build the data layer underneath them, so answers come from actual books, records and rules rather than a model's memory.
A WhatsApp assistant for an Ayurvedic practice: it greets patients in their own language, collects what the doctor needs, and hands the conversation to a human the moment it should. It never interprets symptoms — a deterministic safety gate runs before any model call.
Tamil / English · handoff to practitioner · AWS serverless
All 701 verses with recitation audio, word-by-word meanings and classical commentaries. Every verse is fetched from the corpus and quoted exactly — the model chooses which verse meets the moment, the library supplies the words.
701 verses · audio per verse · Sanskrit + commentaries
A classical-astrology reasoning system built with a practising astrologer. 1,556 scanned pages of Parāśara and Raman, indexed page by page, with source text, corrections and extracted rules kept in separate layers so nothing edits the original.
4 classical texts · 1,556 pages indexed · citation-backed answers
Reads a CV, returns an honest assessment, and runs a structured self-appraisal on any topic the person chooses — with a memory graph that remembers what was assessed and when.
Document understanding · graph memory · tappable questions
Safety rules, state and escalation live in code, not in a prompt. A model can be persuaded; a conditional write cannot.
When a system quotes a source, it fetches that source first. Model knowledge guides where to look, never what to say.
Original text, corrections and derived rules are stored as separate layers. Any interpretation can be revisited without touching the source.
Conversation-level evaluations run against live models before anything reaches a real user's phone.
We take on a small number of projects — conversational assistants, document and corpus pipelines, and AI systems that need to be right rather than merely fluent.