Freelance
AI Chatbots and Assistants
Most chatbot projects stall at the same point: the demo answers well, and then the assistant gives a real customer a confident, wrong number.
Saumya Jain builds AI chatbots and assistants that answer from your own data, with the arithmetic done in code and the model's output checked before anyone relies on it. Based in Jaipur, India, and working remotely.
TypeScript / Python / NestJS / FastAPI / LangGraph / Claude / Amazon Bedrock / pgvector
Start a projectWhat gets built
Assistants over your own documents
Retrieval-augmented assistants that search your documents with hybrid vector and keyword retrieval and cite what they used. In PortfolioGPT, retrieval is a tool the model calls when it needs prose from a document, not a step forced before every answer.
Figures computed in code, not by the model
When an answer involves money or measurements, typed tools compute the figure and the model explains it. PortfolioGPT computes XIRR and asset allocation in TypeScript, and TempoForge's coach can only propose a plan change, which the training engine then checks.
Voice and real-time conversation
Live voice interfaces with streaming transcription and speech, where a dropped connection resumes the session instead of restarting it. Roundwise runs voice or text mock interviews this way.
Guardrails and evaluation
Untrusted text is kept apart from instructions, and model output is validated against a schema with one repair attempt, as in Groundwork. In PortfolioGPT, an evaluation set fails the build when accuracy drops.