Paisa
July 2026 — present
AI CFO & financial operating system
Most finance assistants will happily invent a number. Paisa can't — the
language model never touches arithmetic. A tested engine computes every figure (expenses,
UPI, SIPs, loans, investments, tax) and the model's only job is to explain what the engine
returned. Above that sit four agents with separate remits — accounting, tax, investing,
planning — and a router that decides which one should answer. Bank accounts connect
through secure connectors, so the numbers being explained are real ones.
Built withPython · FastAPI · React Native · TypeScript · AWS · Docker · Kubernetes
Repository↗
NutriScan AI
2026
AI health agent — clinical reasoning, in production
An AI health agent that reasons over symptoms, lab results and clinical
history. Five specialists — doctor, labs, fitness, nutrition and coach — sit behind a
supervisor that routes each question to whichever one should own it, and each specialist only sees the slice of the patient record
it needs. Underneath the chat is the part that matters: a deterministic triage engine of
forty clinical rules that runs before the model reasons, fails closed on error,
and can end the turn outright. A model may raise suspicion; nothing can lower a
rule-derived verdict. Insights carry their own certainty, derived from sample size rather
than asserted, and any number the agent quotes is computed in tested code rather than
produced by the model. 156 tests.
Built withTypeScript · Next.js · Anthropic & OpenAI APIs · Python · FastAPI · PyTorch · PostgreSQL · React Native · Docker
nutritiscan.com↗
Repository↗
Agent framework
2025
Multi-agent orchestration, pulled out of Paisa
The scaffolding underneath Paisa's agents, extracted to stand on its own:
role-scoped agents, tool calling, an intent router that maps a question to the agent that
should own it, and outputs strict enough that the rest of the system can depend on them
instead of re-parsing prose.
Built withPython · LangChain · tool calling · JSON schema
RAG financial assistant
2025
Question answering over statements and filings
Ingestion and chunking tuned for tables rather than paragraphs, embeddings
into a vector store, and answers that cite the row they came from. The citation is the
whole point — an unsourced number is worse than no answer.
Built withPython · embeddings · vector search · LangChain · PostgreSQL