I build end-to-end AI/ML systems — from raw data pipelines to production-ready inference APIs. My work spans classical ML, deep learning, NLP, multimodal models, RAG pipelines, and agentic AI. I care about the full stack: not just model accuracy, but experiment tracking, deployment, and explainability.

Based in Nashik, India · Open to ML/Data roles

ML & AI PyTorch TensorFlow/Keras scikit-learn HuggingFace Transformers CLIP BERT Mistral 7B Phi-3

MLOps & Infra MLflow FastAPI Docker Apache Kafka SQLite ChromaDB

RAG & Agents LangChain Ollama Google ADK Gemini Streamlit Gradio

Data Pandas NumPy Matplotlib Seaborn Plotly SQL

Cloud & Certs Microsoft Azure Oracle Cloud (OCI) Google Advanced Data Analytics

BERT + CLIP gated fusion model trained on MVSA-Single. EDA revealed systematic image-label misalignment (53.3% agreement) — model correctly learned to down-weight the image branch. PyTorch BERT CLIP Gated Fusion Gradio · Macro F1: 0.7084

Fully local RAG pipeline — 1,256 pages ingested into 8,568 chunks, retrieved via ChromaDB, answered by Mistral 7B running on-device. Zero API costs, source attribution on every response. LangChain ChromaDB Ollama Mistral 7B Streamlit · Zero inference cost

4 models tracked with MLflow, alias-based champion promotion, served via FastAPI REST endpoint. PR-AUC prioritized over AUC due to 7% class imbalance. MLflow GradientBoosting FastAPI Pydantic · AUC: 0.868 · PR-AUC: 0.397

Kafka 4.2 (KRaft, no ZooKeeper) streaming GBM-simulated stock ticks at 5/sec. Z-score anomaly detection with rolling 20-tick window, live Plotly dashboard auto-refreshing every 2s. Apache Kafka Streamlit Plotly SQLite · 5 ticks/sec · 2.5σ threshold

Agentic AI assistant built with Google ADK and Gemini 2.5 Flash Lite. Tool-use for product lookup, order tracking, and localized return policies. Eval suite: tool trajectory accuracy 1.0. Google ADK Gemini Tool Use Vertex AI · Trajectory accuracy: 1.0

Fine-tuned Phi-3 Mini (3.8B) using QLoRA on the Spider benchmark (7,000 examples, 160+ databases), training only 0.65% of parameters via LoRA adapters. Improved execution accuracy from 57.83% (zero-shot) to 63.54% — measured by running predicted vs. gold SQL against live databases. Trained end-to-end on a consumer 8GB GPU with 4-bit NF4 quantization and gradient checkpointing. HuggingFace PEFT QLoRA bitsandbytes TRL SFTTrainer Spider · +5.71 pts execution accuracy

LLMOps          ████████░░░░  Eval pipelines → LangSmith → production deployment
Data Engineering ██████░░░░░░  Kafka → Spark → Airflow
System Design   █████░░░░░░░  Scaling ML systems for interviews
CV Fine-tuning  ████░░░░░░░░  Unfreeze CLIP, detection, segmentation