About the developer

Shourya Tiwari

Full-stack engineer with a focus on applied AI systems — retrieval, agents, evaluation, and the infrastructure to keep them honest.

8

delivery phases

60+

engineering-journal entries

400+

backend tests

40+

API endpoints

5

research prototypes

Skills

Languages

TypeScriptPythonSQLCypher

Frontend

Next.jsReactTailwindTanStack QueryReact Flow

Backend

FastAPISQLAlchemyPydanticLangGraphDBOS

ML / AI

Model routingRAGNLI verificationGLiNERLoRA fine-tuningMLflow / DVC

Data

PostgresRedisMemgraphKùzuDBFAISS

Infra

DockerKustomize / HelmZarfGitHub ActionsOpenTelemetry

What this project taught me

  • • Provider-agnosticism is an architecture decision, not a config flag — it has to be enforced (an import-linter contract) or it rots.
  • • A verifier that can fail is worth more than a model that’s usually right. Faithfulness checking changed what the product could honestly claim.
  • • Eval gates make “we chose not to ship this model” a CI-enforced invariant instead of a footnote.
  • • Most “blocked” items weren’t — re-checking the actual environment beat assuming.