portfolio · linkedin · x · pavanpatela5598@gmail.com

Bad systems fail loudly. Mine would fail quietly — a scam gets through, a benchmark scores an agent that did nothing. So I build the thing, then I build the thing that catches it lying.

Evals. 22 container-isolated benchmark tasks, authored. Harbor + Docker, digest-pinned, with verifiers built to be un-gameable — nothing counts until the oracle scores 1.00, an agent that does nothing scores 0.00, and a solution I sabotaged myself also scores 0.00. → a broken task, taken apart

Fraud interception. Rules first on device, retrieval second, model last — so a verdict is explained from evidence, never invented. 12 languages, Android + web, RLS-isolated history. → raksha-ai · private for now, happy to walk through it

Agent safety. Seven agents, none of them in the path between a user and a verdict. Typed contracts instead of chatter, a human gate in front of anything users see, a kill switch and a budget on every loop.

  languages   Python · TypeScript · JavaScript · Rust · Kotlin · SQL · Solidity
  backend     FastAPI · Node.js · Celery · Redis · PostgreSQL · pgvector · Supabase · DuckDB
  frontend    React 19 · Tailwind · Vite · Android
  ml          TensorFlow CNNs · scikit-learn · sentence-transformers · librosa · faster-whisper
  ai          RAG over pgvector · local embeddings · Groq · Gemini · on-box transcription
  agentic     Claude Code · Codex · Cursor · Copilot · Gemini CLI · MCP servers · custom skills
  evals       Harbor · Docker · pytest · NumPy · SciPy
  ops         GitHub Actions · Docker Compose · R8 · signed release builds

_{B.E. Computer Engineering · B.S. Computer Science, BITS Pilani — two undergraduate degrees, in parallel.}