Devraj Gadhvi — AI Engineer. Every project below has a live demo, public source, a story, and proof measured from the build. Nothing here is claimed without evidence.
Story: Enterprise RAG leaks — poisoned documents, broken authorization, confident hallucinations. I built a 5-layer security pipeline: ingestion scanner, authorization enforced inside the SQL query before the model sees data, evidence-grounded synthesis, PII shield, and a hash-chained audit log.
Story: Keyword search fails when you ask "where do we retry failed payments?". This engine maps natural language to code with dense + sparse + RRF fused ranking, explained per result.
Story: Most RAG demos look good once and fall apart on the second question. This one is built to be held accountable: hybrid retrieval (dense + BM25 + RRF), cross-encoder reranking, and every answer chained to verbatim source quotes — unsupported claims return insufficient_evidence. A McNemar eval regression gate fails CI when quality drops.
Story: A multi-tenant AI-employee platform: six n8n workflows (support, lead capture, booking, knowledge ingestion) orchestrated by a Master Orchestrator with LiteLLM intent routing. Isolation is enforced in Postgres Row-Level Security, webhooks are idempotent, and a synthetic replay harness lets anyone docker compose up and watch it run — no real API keys required.
Story: You can't claim safety without testing it. A regression-gating harness for grounded QA: six rule graders, unbiased pass@k and pass^k, cluster bootstrap CI, and an exact McNemar gate that fails CI on a real drop.
Story: A sales-ops agent that scores gone-quiet leads with per-signal attribution, drafts the follow-up, then stops for a named human approval written into a hash chain before anything can send — Postgres RLS isolation, AES-256-GCM vault, idempotent webhooks.
Story: A Model Context Protocol server written against the spec without an SDK — db_query, db_schema, db_tables, db_insert, all with parameterized statements and read-only enforcement on queries. One stdio endpoint that turns any MCP-compatible agent into a safe database assistant.
Story: Research agents should plan, spend, and stop. This orchestrator grows a search tree under a budget with a fixed seed — deterministic, replayable, honest that experiment metrics are simulated.
Story: A pipeline DAG where the served endpoint records which weights it runs and the model card carries the eval hash — 23 tests, stage metrics labelled as deterministic simulation.
hello@dev4aibots.com · github.com/dev4aibots · linkedin.com/in/devraj0770 · Immediate start · Remote worldwide