AI engineering portfolio · Shekhar Dudi
Production-grade AI,
not another demo reel.
Three live systems you can poke right now — hybrid search over millions of profiles, an HR copilot with human-in-the-loop approvals, and a multi-agent content generator. Real backends, real data, real failure modes. The same shape of work I’ve been shipping for 13+ years, made public.
- Live solutions
- 3
- Open source
- 100%
Solutions
Each card opens a working demo backed by a FastAPI service. Architecture diagrams and API contracts are documented inline — so you can read the code, hit the endpoint, or just play with the UI.
Intelli Search
Hybrid search over millions of company profiles — an intent classifier routes each query through BM25, vector kNN, or an agentic web-augmented pipeline.
- LLM intent classifier picks regular / semantic / agentic per query
- BM25 + vector kNN fused with Reciprocal Rank Fusion
- Agentic mode plans, web-searches, and rewrites for off-corpus questions
- Redis cache, circuit breakers, PII scrubbing on the hot path
Agentic HR
LangGraph HR copilot that triages intent into policy RAG, leave operations, or access provisioning — with manager approvals before anything mutates.
- Intent triage routes to policy / leave / access subgraphs
- Self-grading policy RAG over pgvector with citations
- Human-in-the-loop approval gate on every write
- Approved actions fulfill via Gitea + Mattermost; full audit trail
LinkedIn Auth
Two-stage pipeline — Pulse Scout briefs from the open web, Authority Crew turns a chosen angle into a posted-ready draft and cover image.
- Pulse Scout: Agentic Market Analyst for AI trends, across multiple signals and sources
- Multi-agent Crew: Researcher → Writer → Critic → Visual Director
- Async job API with streaming progress; survives long timeouts
- Per-IP rate limits, cover image via gpt-image-1, configurable voice
How these are different
A “portfolio” of LLM demos is easy. A portfolio that survives contact with users is the actual job. Four things I optimise for — in this site and in production.
Demos are real services
Every card opens a live FastAPI backend — same code, same shape as production. Open the network tab; nothing is mocked.
Human-in-the-loop where it matters
Agents act, humans approve. Guardrails, structured outputs, and audit trails are first-class — not afterthoughts.
Scalability without complexity
Three products, highly scalable backends. Extremely fault-tolerant reusable architecture patterns, not bespoke spaghetti.
Engineered Intelligence
Agentic where it needs to be, engineered everywhere else. No black boxes, no magic prompts — just solid systems.
Built with
Same stack, scaled downWorking on something agentic?
I help teams move ai and agentic systems from clever demos to billable systems. Always happy to compare notes, sanity-check architecture, or talk roles.