Work · 10 projects

Projects I’ve built.

Chosen for technical depth, measurable results and end-to-end product thinking. Open any project for the problem, how it works and the decisions behind it.

Agent infra Currently building

SHAPI

Self-maintaining APIs for teams and AI agents.

  • Maps every third-party call site into an api.lock
  • Fixes kept only when a contract test proves them
  • Opens a pull request with the fix, test and evidence
  • TypeScript
  • Node
  • MCP
  • Claude
Agent product

Nori

Job search infrastructure built around a career graph.

  • Career graph: every resume bullet traces to a fact
  • Fills ATS forms with review-before-submit
  • Drive it from Slack, Telegram, WhatsApp, email or MCP
  • TypeScript
  • Agents
  • MCP
  • Composio
  • Parallel
Multilingual AI

MotherTongue

Translate text, photos and documents into the language you understand best.

  • Text, photo or document in; your language and audio out
  • 150+ language interface, installable PWA
  • Keys stay server-side; routes are rate-limited
  • Next.js
  • TypeScript
  • OCR
  • Speech
Agent product

Chief

A personal agent that notices, proposes, and acts only after you approve.

  • Notices problems across calendar, email and sleep
  • Every write waits for an Approve tap
  • New skills are single files of about 20 lines
  • Python
  • Agents
  • Composio
Agent infra

Kerna

A local-first trust layer for AI agents.

  • Fail-closed permissions for every MCP tool
  • Hard budgets on calls, runtime, cost and output
  • A local receipt for every agent run
  • Rust
  • MCP
  • SQLite
Evaluation

RAGTruth

Can multiple AI agents detect when RAG systems hallucinate?

  • Overall F1 of 0.642 with no training data and no GPU
  • High recall of 0.764 on claim-level hallucinations
  • Task-aware rules drive the largest gain (+0.13 F1)
  • Python
  • LLMs
  • Multi-agent
Computer vision

Video Search

Semantic search across video scenes using natural language.

  • Automatic scene detection with CLIP embeddings
  • Natural-language search over scenes in Qdrant
  • Full-stack app with a FastAPI backend and React frontend
  • CLIP
  • Qdrant
  • FastAPI
Agent infra

AgentOps

A control tower for tracing, debugging and auditing agent runs.

  • Every agent run recorded as a replayable trace
  • Risk classification and human approval for risky steps
  • Demo agent writes a safe runbook for a failed API incident
  • FastAPI
  • React
  • PostgreSQL
  • SQLAlchemy
ML systems

Feature Platform

Real-time feature computation and multi-model serving.

  • Kafka → Feast → Redis streaming feature pipeline
  • Fraud, recommendation and forecasting models share features
  • A/B testing with Ray Serve, monitoring with MLflow
  • Kafka
  • Feast
  • Redis
  • Ray Serve
  • MLflow
Document AI

Document Intelligence

A multi-agent system that reads, extracts and answers questions over business documents.

  • 82% accuracy on a 150-document test set, under 5 s per document
  • 85% top-3 relevance across 50 semantic search queries
  • 84% reduction in manual entry time
  • LangChain
  • Qdrant
  • FastAPI
  • RAG
  • Python

Archive

More from GitHub.

Smaller experiments, course work and companion repos. These link straight to the source.