I build reliable AI products and backend systems end to end—from retrieval-grounded applications and governed agentic workflows to AI-native full-stack products and applied ML systems.
My work focuses on turning AI capabilities into usable software with clear boundaries, evaluation, observability, and human control where it matters.
- Agentic systems with memory, planning, tool use, bounded autonomy, and human approval
- Enterprise RAG with authorization-aware retrieval, hybrid search, PII protection, and evaluation
- AI infrastructure for model routing, local/cloud inference, cost control, and scale
- AI observability for traces, tool calls, hallucination signals, loops, latency, and cost
- AI-native products across collaboration, commerce, software engineering, and operations
- Multilingual and voice AI designed for real users and real-world constraints
| Project | What it demonstrates |
|---|---|
| Secure Enterprise RAG | Multi-tenant, authorization-aware RAG with JWT, RBAC, PII protection, PostgreSQL/pgvector, hybrid retrieval, and security evaluation. |
| OWNARA AI | Governed AI execution with bounded authority, human approval, measurable outcomes, and auditable workflows. |
| Repo Pilot | Evidence-based software engineering intelligence for codebase understanding, diagnosis, code review, and developer assistance. |
| Sahayak | AI-assisted meeting coordination for Hindi, English, and Hinglish teams. |
| Bhashini Voice Gateway | Multilingual voice-to-action workflows with ASR, LangGraph intent routing, human confirmation, and safe actions. |
| Demand-Risk Intelligence Engine | Applied ML system using seeded enterprise telemetry for demand forecasting, inventory risk scoring, and explainable replenishment recommendations. |
AI/ML: LLM applications · RAG · agents · LangGraph · embeddings · evaluation · prompt engineering · PyTorch · LoRA · speech AI
Backend & data: Python · FastAPI · TypeScript · Next.js · PostgreSQL · pgvector · Redis · RabbitMQ
Production: Docker · AWS · Kubernetes/EKS · authentication · RBAC · multi-tenancy · observability · CI/CD
- Reliability over demos: evaluate behavior, handle failure modes, and make systems observable.
- Security by design: enforce authorization at retrieval and action boundaries, not just at the UI.
- Human control where it matters: use bounded autonomy, approvals, and audit trails for consequential actions.
- Product thinking: optimize for user outcomes, latency, cost, and maintainability—not model novelty alone.
Reliable long-running agents · evaluation-driven development · efficient model routing · secure AI infrastructure · multilingual AI products
Open to conversations about applied AI, AI platform engineering, and building useful products with modern AI systems.