I'm an AI/ML engineer who builds LLM applications end to end: RAG pipelines, fine-tuned models and multi-step agents, shipped with Docker and AWS. I also teach the Full Stack AI Engineer course at DrippinAI Labs.
const punit = {
role: "AI/ML Engineer",
languages: ["Python", "JavaScript", "SQL"],
focus: ["Agentic AI", "RAG", "LLM Fine-tuning", "Generative AI"],
agents: ["LangGraph", "LangChain", "CrewAI", "n8n"],
retrieval: ["FAISS", "Pinecone"],
models: ["LLaMA", "Hugging Face Transformers", "Groq"],
classicalML: ["scikit-learn", "TensorFlow", "Keras", "PyTorch"],
deploy: ["Docker", "AWS", "Streamlit"],
currentFocus: "Production-grade agentic AI systems",
funFact: "There are two ways to write error-free programs; only the third one works"
};- π Ex-AI Intern at Industrial Light & Magic (ILM), Production Technology, Mumbai
- π AI Lead Instructor at DrippinAI Labs, Full Stack AI Engineer course
- π± Going deeper on agent orchestration, evaluation and LLM fine-tuning
- π¬ Ask me about RAG, agentic AI, LLMs, Python and ML
- π« Contact: punitgavali29@gmail.com
I design agents as explicit state machines (LangGraph), not one giant prompt. Example flow from my recruitment pipeline:
flowchart TD
A[Resume + Job Description] --> B[Shield: injection and hidden-text check]
B --> C[Blind redaction: names and schools hidden]
C --> D[Scorer A]
C --> E[Scorer B]
D --> F[Quote verifier: no quote, no credit]
E --> F
F --> G[Adjudicator]
G --> H[Decision policy]
H --> I{Recruiter approves?}
I -- Yes --> J[n8n: Discord alert + candidate email]
I -- No --> K[Held for review]
| Project | What it does | Stack |
|---|---|---|
| π Meridian: Evidence-Grounded Multimodal RAG | Multimodal RAG for industrial maintenance documents with 8 retrieval channels, page citations, conflict resolution, claim verification and evidence-based abstention. Achieved 90% answer correctness, 100% Hit@5 and 100% abstention recall on held-out evaluation. | Python, Qdrant, Docker Compose, Ollama |
| π§βπΌ HirePanel: Multi-Agent Recruitment Pipeline | Multi-agent resume screener where every score needs a verified quote. Blocks prompt injection, hides names and schools, and a human approves every decision before n8n sends Discord alerts and emails. | Ollama, FastAPI, n8n, Docker, MCP |
| π’ Enterprise RAG Chatbot | Support assistant that cut support workload by about 42%. | RAG, vector DB, LLMs |
| π₯ MedQueryDoc AI | Medical document Q&A. | LLaMA-2, Pinecone |
| π§Ύ Receiptorium Chatbot | Fine-tuned LLaMA 3.2 assistant for a work-tracking platform. | LLaMA 3.2, fine-tuning |
| π³ Transaction Anomaly Detection | Flags suspicious financial transactions on a live dashboard. | Isolation Forest, Streamlit |
| π‘οΈ PhisProbe | Phishing link detection, published in IJNTI, IIT Roorkee Productathon finalist. | Streamlit, GenAI |
- π― Finalist at Productathon AI Hackathon, IIT Roorkee
- π Published research paper in IJNTI Journal: "PHISPROBE: Phishing Link Detection Using Streamlit and Generative AI Approach"
- π IBM AI Engineering Professional Certificate (2024)
- π B.E. in Artificial Intelligence & Data Science, FR CRCE Mumbai (8.82 CGPA)






