My Builds
10+
End-to-End
Projects Owned
Technologies Used: UVM, SystemVerilog, Verilator, LangGraph, Agents, RAG, MCP
Built a 7-agent LangGraph system that plans, writes, and runs SystemVerilog/UVM tests for RTL chip designs, end to end, automatically
Retrieves cited passages from DUT spec sheets via hybrid RAG (BM25 + bge-small embeddings, RRF) to ground the generated test plan
Verified a dual-port RAM via Verilator, closing 100% functional coverage in 43 cycles of directed stimulus over 3 iterations
Validated AI-written testbenches with mutation testing (stuck-at, inversion, swap, off-by-one): 10 of 10 planted bugs caught overall
Exposed the whole agent toolchain as an MCP server, so Claude Code or any other AI client can write tests and verify a RTL directly
Technologies Used: LLM Fine-Tuning, LoRA, QLoRA, DeepSpeed, vLLM, MLX
Fine-tuned Qwen2.5-Coder (1.5B-7B) across 19 LoRA/QLoRA/full runs with DeepSpeed ZeRO-2 on H100/H200 GPUs, via W&B
Built data pipeline over 131K OpenRTLSet Verilog modules, removing 54K fuzzer-generated rows and holding out a compiling test set
Evaluated via Verilator differential testing and VerilogEval, raising pass@1 from 17.3% to 44.9%, beating a 9x larger 14B base model
Served models via auto-detected MLX/vLLM/Ollama backends (Streamlit, CLI, VS Code); 4-bit quantization gave 3.5x faster decode
Technologies Used: RAG, LangChain, Gemini APIs, Explainable AI, Guardrails, CI/CD
Integrated Gemini 2.5 Pro (VertexAI) with RAG pipeline (FAISS + HF embedding) to handle 20K+ daily chats with chat persistence
Architected Full-stack application (TypeScript + FastAPI), with SSE streaming, JWT auth, and automated CI/CD on Cloud Run
Encrypted user data with AES-256 and rotating keys (GDPR compliant) in Firebase while maintaining <1s p99 TTFT for 600+ DAU
Technologies Used: Distributed Systems/Training, CUDA, Fine-tuning, vLLM, Ollama
Fine-tuned Flux.1-dev on 1,646 WhatsApp emojis using LoRA with DeepSpeed Zero-2 on 2 Nvidia A100 GPUs, cutting model loss to 0.245
Designed Prompt Assist (2-shot Prompting + Llama 3.3) to refine user inputs before fine-tuned model, improving emoji quality
Boosted performance by 40% via GPU-accelerated inference (vLLM/MLX + Diffusers/Mflux + Streamlit) and Int8 Quantization
Technologies Used: HuggingFace, OpenAI APIs, Microservices, LLM Optimization, BigData
Architected a Chatbot with an end-to-end multi-agent RAG leveraging 4 LLMs using LangGraph, Guardrails, FAISS, and BigQuery
Built the backend with FastAPI, deployed on Google Cloud Run (Docker + Git Actions CI/CD), capable of handling 20k queries/sec
Evaluated with DeepEval, achieving Correctness 0.81 and Answer Relevancy 0.87 for high-quality, relevant responses
Tracked experiments with MLflow and monitored traces with LangSmith, improving model performance via human-feedback loops
Optimized state management via layered Redis caching, enabling 3× faster responses (1.5s p95 E2E and 250 ms p95 TTFT) under 6× load
End-to-End Kidney Tumor Classification
Technologies Used: CNN, Clinical Data, Compute Engine, Docker, Airflow, CI/CD
Built an end-to-end image classification pipeline using a modified VGG16 CNN model using PyTorch, achieving an F1 score of 0.96
Developed interpretable ML workflows with feature importance analysis, SHAP-based insights, and calibration for bias reduction
Automated data ingestion, model training, and evaluation pipelines using Airflow, MLflow, and DVC for reproducibility and tracking
Containerized and deployed the trained model as a Flask web app on AWS EC2 via Docker and Git Actions for real-time inference
Fantasy Team Recommendation for IPL 2024
Technologies Used: Fine-tuning, QLoRA, Prompt Engineering, NLTK, Transformers
Fine-Tuned Gemma 2 using QLoRA on cricket dataset resulting in 10% improvement in ROUGE score for cricket-specific text generation tasks
Established algorithm using NLTK to extract structured data from unstructured IPL historical data including player stats and match scorecards
Leveraged Prompt engineering (2-shot) to enhance prediction accuracy resulting in 85% accurate team prediction for the IPL matches
Rent the Runway Fashion Recommender System
Technologies Used: Collaborative filtering, SVD, Cosine Matrix, Web Scraping
Scraped product data and user reviews via BeautifulSoup, creating structured datasets with detailed attributes and feedback metrics
Engineered preprocessing pipelines for text cleaning, tokenization, and feature extraction to enhance review relevance
Developed hybrid recommendation using matrix factorization for collaborative and cosine similarity for content-based filtering
Incorporated incremental SVD updates to address cold-start users and products, cutting system update time by 40%
Advanced Performance Metrics for Ultimate Frisbee Athletes
Technologies Used: Predictive Modeling, Bayesian analysis, ML Pipelines
Engineered player rating models by combining XGBoost and mixed-effects linear modeling for context-aware performance evaluation
Designed on/off plus-minus models with Bayesian modeling, increasing predictive accuracy 20% for offensive and defensive impact
Developed composite scoring metrics integrating multiple player features, improving ranking robustness by 15%
Built reproducible pipelines for model training and validation using DVC and MLFlow, ensuring scalability and consistent evaluation
End-to-End SQL Database Chatbot
Technologies Used: Agent, ORM, SQL, LangChain, Groq
Developed an AI Agent using Langchain, SQL, and Gemma 2, allowing users to query databases with Natural Language Queries
Leveraged SQLAlchemy ORM and custom connection handling to support both SQLite and MySQL databases dynamically
Created SQL Database Toolkit with Gemma 2 zero-shot question answering, enhancing the chatbot's ability to handle complex SQL queries accurately