██████╗ █████╗ ██╗██████╗ ██╗ ██╗████████╗ ████████╗
██╔══██╗██╔══██╗ ██║██╔══██╗██║ ██║╚══██╔══╝ ╚══██╔══╝
██████╔╝███████║ ██║██████╔╝██║ ██║ ██║ ██║
██╔══██╗██╔══██║██ ██║██╔═══╝ ██║ ██║ ██║ ██║
██║ ██║██║ ██║╚█████╔╝██║ ╚██████╔╝ ██║ ██║
╚═╝ ╚═╝╚═╝ ╚═╝ ╚════╝ ╚═╝ ╚═════╝ ╚═╝ ╚═╝
Building systems that think, retrieve, stream, and adapt.
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
github.com/rajput-t/llm-finetuning-text2sql(in progress)
Fine-tuning Phi-3 Mini (3.8B) on the Spider benchmark (7,000+ Text-to-SQL examples) using QLoRA — 4-bit quantization via bitsandbytes, LoRA adapters via PEFT, training on RTX 2070 Super (8GB VRAM).
Target: 78%+ execution accuracy, outperforming zero-shot baseline and benchmarking against GPT-3.5 zero-shot.
HuggingFace PEFT QLoRA bitsandbytes TRL SFTTrainer Spider
| Repo | Description |
|---|---|
deep-learning-fundamentals |
ANN, CNN (CIFAR-10), LSTM — early DL experiments in Keras/TensorFlow |
ML_algorithms |
Linear Regression, Decision Trees, Random Forests, KNN, K-Means, PCA |
leetcode |
DSA problem solving — ongoing |
certificates |
Google · Microsoft Azure · Oracle Cloud · BCG Forage |
LLMOps ████████░░░░ Fine-tuning → eval pipelines → deployment
Data Engineering ██████░░░░░░ Kafka → Spark → Airflow
System Design █████░░░░░░░ Scaling ML systems for interviews
CV Fine-tuning ████░░░░░░░░ Unfreeze CLIP, detection, segmentation
"Build systems, not demos."