|
I'm Saksham Pathak, a Generative AI Engineer and M.Sc. AI & ML graduate of IIIT Lucknow, building intelligent systems that can listen, reason, retrieve knowledge, and take action. My work spans Voice AI, Agentic Systems, RAG Pipelines, Fine-Tuning, and LLM Testing, with a strong interest in real-time AI experiences that feel natural and human. Currently, I'm exploring:
I enjoy turning ambitious AI ideas into practical systems that people can interact with, trust, and use every day. |
|
| Domain | Stack |
|---|---|
| Voice AI | Whisper · Deepgram · WebRTC · Custom VAD · STT / TTS · Twilio |
| Agents | Google ADK · LangGraph · LangChain · AutoGen · MCP · Tool Calling |
| LLMs & Training | PyTorch · Transformers · LoRA · PEFT · Gemini · OpenAI · Claude · Llama |
| RAG & Vector | FAISS · ChromaDB · Pinecone · Hybrid Search · Reranking |
| Eval & Safety | ghostrun · LLM-as-a-Judge · Ragas · Guardrails · PII Redaction |
| Infra | FastAPI · WebSockets · Redis · MongoDB · Docker · GCP Cloud Run · Vercel |
01 — ghostrun · Testing for LLM Apps
pytest-native record/replay for LLM calls, semantic assertions via a local judge, and Bayesian prompt optimization. Offline, free, deterministic CI.
Python pytest Ollama · 
02 — Saarthi AI · Commute-Planning Agent
Autonomous Google ADK agent that sweeps future departure windows and fuses traffic, weather and local events into a transparent risk score, with persistent MongoDB MCP memory.
Google ADK Gemini MongoDB MCP FastAPI · 
03 — Real-Time 3D Conversational Avatar
Speech → LLM → TTS with barge-in, synchronized facial blendshapes and lip-sync.
PyTorch WebRTC TTS Custom VAD ·
04 — DocuMind AI · Document Intelligence
RAG-based Q&A and summarization over PDFs with OCR, FAISS retrieval and an MCP server on Gemini.
FastAPI FAISS Redis Gemini ·
05 — parth-dl · Open Source CLI
Zero-dependency Instagram media downloader. No login, no API key. Published on PyPI.
Python CLI ·
06 — 100 Best GenAI Projects 2025
Curated 100+ high-impact open-source LLM, RAG & Agent projects.
LLM RAG Agents ·
- GitHub Docs — Merged PR #45118: fixed a duplicated word in image alt text for the Copilot access docs.
"The goal isn't to build models. The goal is to build intelligence that matters."