Hi, Ranjan here. I build AI agents, RAG systems, and voice pipelines. Sometimes full-stack. Always shipping.
| Fillo v2 Multi-Agent · Claude Code Hackathon |
Resume + job URL -> 6 specialized agents (Researcher, Skill Mapper, Writer, Critic, Form Agent, Reviewer) with a Critic feedback loop. Built with CrewAI + Browser-Use + Inngest. Free-tier first via Groq Llama 3.3 + Gemini 2.0 Flash. |
| Artha Voice AI · Finance Agent |
Voice finance agent in Hindi, English, Hinglish. Agent loop from scratch, no LangChain. 92% intent accuracy on a 50-case eval set. Stack: Groq Whisper + Sarvam Bulbul v3 + FastAPI + WebSockets. |
| MnemOS Agentic OS · Hackathon |
Agentic OS that remembers across sessions. Remember / Recall / Recover / Plan nodes over HydraDB graph search. Runs in a containerized virtual desktop (Playwright + noVNC). Supports Groq, Claude, Ollama. |
| WEBGhosting-MCP MCP Server · Go · ⭐ 28 stars |
Stealth browser MCP server for AI agents. 30 tools, 22 anti-fingerprint scripts. Works with Cursor, Claude, VS Code, and any OpenAI-compatible client. Built in Go. |
| DevRadar Full-Stack · WikiThon 2026 |
AI career intelligence platform for Indian devs. Live knowledge graph, wiki ingest pipeline, grounded career chat, learning roadmap. Deployed on Vercel + Render. Stack: React 18 + Node.js + Groq + Claude + HydraDB. |
| Off-Duty Google Cloud Hackathon · MongoDB |
Inventory-aware AI store manager for local cafes. Proactively bundles what the shop needs to move, not just what is on the menu. Stack: Google Cloud Agent Builder + Gemini + MongoDB MCP + FastAPI + React. |
| Document Intelligence RAG · Production-style |
RAG pipeline with hybrid search (BM25 + vector fusion) and semantic chunking. Low-latency via Groq Llama 3.3. Quantified with RAGAS metrics. |
| Messy Mashup Audio ML · Kaggle |
Genre classification from noisy audio mashups. XGBoost -> CNN -> CRNN -> fine-tuned HuBERT. 94.89% F1. Live demo |
| Voice AI Agent Voice · 3-tier intent |
Voice assistant with 3-tier intent classification (Regex -> Keywords -> LLM). Compound commands, session memory, sandboxed tool execution. Groq Whisper + Llama 3.3. Demo |
- How We Classified Music Genres From Noisy Audio -- HuBERT, domain shift, and the data trick that hit 94.89% F1
- The Unsexy Secret -- The 50 Example Method for LLM apps
- Token Context Bottleneck -- Architecting the Future of Agentic Tool Use
- Context Window Problem -- How We Solved It for Large-Scale Web Scraping
|
AI / LLM Voice RAG & Vector ML |
Frontend Backend Infra |


