Navneet Prabhakar

I lead engineering teams at Target while staying deeply hands-on — designing production-grade AI agent systems, multi-LLM platforms, and enterprise Java. My work lives at the intersection of cutting-edge AI and code that actually ships to production. Most of my recent energy goes into agentic developer tooling: turning natural language into planned, reviewed, verified software.

A multi-agent system that turns natural-language change requests into plan-gated, verified pull requests across many repos and projects. A Spring Boot 3 + Spring AI engine drives a multi-LLM router (Anthropic · OpenAI · Gemini) with per-provider failover and cost tracking, RAG over project docs (pgvector), and SWE-bench-style empirical evaluation. It ships as a full product surface — a TypeScript CLI (installable via Homebrew), a VS Code extension, live preview infrastructure (auto-deployed through GitHub Actions), and a docs site. It grew out of my earlier DevPilot prototype: brainstorm → plan → implement → AI review, with plan versioning, digital sign-off, event-sourced audit trails, and OAuth2/JWT.

An orchestration layer on top of Claude Code: builder agents work in isolated domains, an adversarial reviewer agent approves or rejects every PR, and the whole crew is driven from Slack. Approved PRs auto-merge, rejections loop back with feedback, repeated failures escalate to me — with a morning digest. Wake up to merged PRs.

A local-first training workspace for Gemma 4 12B with MLX-LM + LoRA on Apple Silicon — reproducible gate-based experiments, holdout-loss tracking, and quantized (q4/q8) serving paths. Proof-on-laptop before any GPU/cloud spend.

A candidate-first job portal for India + US where an LLM agent is the spine, not a feature. "Today's Brief" pre-warms curated matches each morning; a hybrid keyword + LLM tool-router drives 8 tools (search, tailor resume, research company, draft counter-offer, career advice…). Weighted match scoring, H-1B / LCA sponsorship signals, salary intelligence with regional take-home calculators, and a reskill planner — all feature-flag-gated.

An AI prep platform for Indian exams: AI tutor with a RAG abstraction, practice/tests, mistake notebook, progress tracking, safeguarding, and parent boundaries. Next.js + Fastify + PostgreSQL/pgvector monorepo.

• project-chaos — Multi-LLM debate orchestrator: concurrent OpenAI / Anthropic / Gemini agents with a supervisor watching consensus in real time. Spring Boot 3.4 + Spring AI + React + STOMP/WebSocket
• trade-mcp-server — Stock-trading MCP server built on Spring AI (Groww)
• mcp-client · mcp-server — Reference Model Context Protocol client / server
• zero-trade-app — Agentic trading on Zerodha Kite Connect: technical / news / research agents + risk manager
• fno-conservative-algo — Conservative F&O algo trading with AI agents

• openai — OpenAI + Spring AI with RAG, pgvector, chat memory
• anthropic — Claude integration with Spring AI
• gemini — Gemini API with Spring AI
• javagpt — Lightweight local GPT in Java 21 + Deep Java Library (DJL)
• query-builder — Text-to-query / text-to-SQL generator

• telegram-bot-llm — Telegram bot with LLM code-gen
• telegram-bot — Telegram bot with LLM + MCP capabilities
• websocket — Simple chat over WebSockets

• drools ⭐ 5 — Drools Rule Engine + Spring Boot
• executor ⭐ 3 — Java Executor Framework patterns
• credit-risk-analyser — Credit risk from bank statement + bureau + GST/ITR, with Drools
• bank-statement-analyser — Bank-statement categorisation (OCR + PDFBox + Gemini)
• kafka — Kafka producer/consumer + Spring Boot
• elasticsearch — Elasticsearch + Spring Boot
• dms — Document Management System (MongoDB GridFS)
• jeasy — Jeasy Rule Engine + Spring Boot + Swagger
• stock-exchange-trade — Groww API wrapper for stock & F&O
• fno-dashboard — Investment & strategy tracking UI
• cowin — CoWin Public API integration

• montecarlo ⭐ 1 — Monte Carlo simulations
• geneticAlgo — Genetic algorithm implementation
• tictactoe — Tic-tac-toe with minimax AI
• network — Mail-train control system with Dijkstra scheduling (Java)

• uxdesigns — Curated DESIGN.md design-system files agents can read to build matching UI

Specialties: AI agent orchestration · multi-LLM routing & failover · Model Context Protocol (MCP) · RAG + pgvector · prompt caching · LLM fine-tuning (LoRA / MLX) · rule engines (Drools, Jeasy) · distributed systems · real-time messaging (Kafka, STOMP, SSE) · observability (Micrometer, OTel, Prometheus, Grafana) · algorithms (Monte Carlo, Genetic, Minimax, Dijkstra)

• Spec-driven, plan-gated codegen workflows that keep the human in the loop
• Cost-aware LLM routing with prompt-cache hit-rate optimisation across Anthropic / OpenAI / Gemini
• Empirical, SWE-bench-style evaluation of agentic systems on real-world bug fixes
• Fine-tuning small coding models locally (Gemma + MLX + LoRA) before scaling to cloud