Navneet Prabhakar

I lead engineering teams at Target and build personal projects in AI developer tooling, coding-model experimentation, and enterprise Java. My open-source work explores how specifications, agent workflows, and executable verification can produce more reliable software changes.

Specifications, agent workflows, code review, and verified software changes — with Spring Boot, Spring AI, and multi-provider LLM integrations.

Active research into predicting the consequences of a software change before execution. The public experimental release includes bounded controller candidates, runtime tooling, and evaluation evidence. The broader engineering goal remains open.

Closed experiments in coding-model adaptation, training-data provenance, and executable evaluation. Full source histories and evaluation materials are public; released checkpoints are linked below. These are experimental records, with results and limitations documented in their repositories and model cards.

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.

Published under ExecuteSpec R&D Lab. See each model card for intended use, formats, evaluation evidence, and limitations.

• 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 — Local GGUF inference in Java with Jlama and llama.cpp
• 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 — Drools Rule Engine + Spring Boot
• executor — 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 — 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
• Software world models, small coding models, and reproducible evaluation with preserved train/test boundaries