I'm a Software Engineer at E2open in Bengaluru, building production ML and GenAI systems. I've spent 3+ years taking models from experiment to production: LLM apps and RAG, fine-tuning, inference optimization and the MLOps around them.
I care most about the engineering around the model: making it fast, measurable and reliable once real users depend on it.
- 🤖 LLM support chatbot that automatically resolves customer support tickets
- 🔎 RAG pipelines on LangChain with FAISS and Chroma, evaluated with Ragas and DeepEval for faithfulness, relevance and context precision
- 🧩 Agentic workflows with LangChain and CrewAI, plus MCP integrations that give LLMs secure access to enterprise databases
- 🎯 LLM fine-tuning with PEFT (LoRA, QLoRA) for domain-specific and multilingual use cases
- ⚡ 38% lower inference latency through quantization and pruning with ONNX Runtime and Hugging Face Optimum
- 📈 9.5% higher accuracy from replacing a legacy text classifier
- 📊 Production monitoring with FastAPI, Airflow, Prometheus and Grafana to catch model performance drops
| Project | What it shows |
|---|---|
| mini-chatgpt | Built a chat assistant from a base LLM: SFT → DPO → benchmarks. Matches Qwen's official instruct model on instruction following (IFEval) and beats it on ARC with ~3 GPU-hours |
| generative-ai-engineering | QLoRA text-to-SQL (7B), 7B → 0.5B distillation, GRPO reinforcement learning, function-calling agent (16.7 % → 94.8 % exact tool calls), RAG evaluation, vision-language, Whisper for Marathi, prompt-injection guardrail |
| deep-learning-portfolio | 14 projects across computer vision, NLP and generative models, e.g. BERT made 4.8× faster via distillation + ONNX + int8, Vision Transformers, U-Net, AdaIN, Music Transformer, molecule generation |
| cnn-edge-optimization-tflite-onnx-tvm | Compressing a Keras CNN for edge devices with TFLite, ONNX and Apache TVM; models ~12x smaller. Final-year team project |
| llm-token-visualizer | Streamlit app comparing how LLM tokenizers (BERT, GPT-2, GPT-4, Phi-3, DeepSeek) split the same text |
| surname-origin-classifier-rnn | Character-level RNN in PyTorch that predicts a surname's language of origin across 18 languages, served with Streamlit |
| vm-energy-agent | Monitoring agent that collects system metrics, estimates VM power usage and exposes them to observability systems |
GenAI: LangChain · LangGraph · CrewAI · Hugging Face · PEFT · FAISS · Chroma · Ragas · DeepEval · MCP
ML: PyTorch · TensorFlow · scikit-learn · ONNX Runtime · Optimum
MLOps and data: FastAPI · Docker · Kubernetes · MLflow · Airflow · Prometheus · Grafana · Spark · Kafka · Snowflake
Cloud: GCP (Vertex AI) · AWS (SageMaker, Bedrock)
🎓 B.Tech in Computer Science and Engineering, DKTE Society's Textile and Engineering Institute, Ichalkaranji (2019–2023)
