AI Engineer building LLM agents, RAG systems, and evaluation pipelines that hold up outside a notebook. B.Tech in AI & Data Science, IIT Jodhpur ('26) · Open to remote full-time and internship roles.
📫 sushrutbarmate@gmail.com · LinkedIn
- LlamaIndex: 5 merged PRs, including a
PostgresChatStoreserialization fix,ReActAgent/ OpenAI integration fixes, and retriever improvements, all reviewed by core maintainers. - Stagewise: fixed an Electron update/restart crash (
TypeError: Object has been destroyed) by replacing synchronous RPC calls with fire-and-forget procedures. Merged into main.
AI/ML Engineering Intern · LeMiCi IQ (Remote, Sep–Dec 2025) Built multi-agent RAG pipelines (LangChain, ChromaDB, Gemini, FastAPI) with sub-2s end-to-end latency, and improved retrieval accuracy by 24% through dynamic chunking. Built an ingestion pipeline that processed 70,000+ records.
ForgePilot: a constrained coding agent
Turns natural-language goals into tested Python projects: generate → validate with pytest → self-heal from tracebacks → fall back to last-known-good. Hard LLM-call budget, exact + semantic caching, and predictable cost per run.
StreamClip AI: Create YT Shorts from Livestreams
Turns a livestream VOD into vertical, captioned comedy clips for TikTok, YouTube Shorts and Reels. Paste a Twitch/YouTube/Kick VOD link, and it finds the funny moments, cuts ~30 second clips around them, and renders them 9:16 with the facecam on top, gameplay below, animated captions, a hook, and meme sound effects.
LLM Code Generation Benchmark: zero-shot vs RAG
Evaluated DeepSeek-Coder, Qwen-Coder, LLaMA 3.1, and Gemini 2.5 Flash using functional tests plus Radon, Bandit, and COCOMO metrics. Finding: zero-shot achieved 16% higher functional test coverage than RAG on this benchmark, because retrieval overhead can hurt direct code generation.
AI: LLMs · RAG · AI Agents · LLM Evaluation · LlamaIndex · LangChain · HuggingFace · PyTorch Engineering: Python · FastAPI · PostgreSQL · Docker · GCP · Pytest · Git · Linux