Machine Learning Engineer · AI Systems · Reinforcement Learning · NLP
Computer Science Undergraduate at IIIT Kottayam · India

I build AI/ML systems that ship — focused on Reinforcement Learning environments, multi-agent LLM orchestration, and retrieval-augmented systems.

• 🎯 Grand Finalist at the Meta × Scaler OpenEnv Hackathon (52,000+ participants)
• 🤖 Currently building multi-agent RL environments where small LLMs learn to route work across specialists
• 🔍 Past work spans cross-lingual RAG (English / Hindi / Kannada / Malayalam), agent-based system design, and ESG analytics
• 🏆 2nd Runner-up at Hackzilla National Hackathon (IIIT Sonepat)
• 📚 200+ DSA problems solved · CGPA 7.8

Languages

AI / ML

LLM / NLP / Retrieval

Backend & Deployment

Cloud & Tools

Multi-agent OpenEnv environment where a Chief of Staff routes work across four AI specialists — Data Analyst, Finance, Strategy, HR — to complete CEO briefing tasks. GRPO-trained Qwen2.5-1.5B policy with optional RAG over company memory.

My contributions: HR/Communications subenvironment with blended grader (45% structure + 45% tone + 10% audience), deployment to HF Spaces (Docker + Git LFS), validation infrastructure (live HTTP smoke tests, GitHub Actions CI, session capping), README + technical writeup.

🔗 Live Demo · Blog

Meta × Scaler OpenEnv Hackathon — Grand Finalist (Round 2)

Developed a multi-task computer vision pipeline for aerial drone imagery that simultaneously predicts pixel-precise Ground Control Point (GCP) coordinates and marker shape classification using a ResNet-50 backbone.

Highlights

• 🎯 Built a multi-task architecture combining heatmap-based keypoint localization with shape classification (Cross, Square, L-Shaped).
• 📈 Improved localization by replacing direct coordinate regression with Gaussian heatmaps + soft-argmax decoding, achieving over 40× improvement in PCK@25 compared to the regression baseline.
• 🧠 Identified and fixed a zero-gradient training bug by introducing joint optimization with MSE + CrossEntropy loss, achieving 100% validation Macro F1 for shape classification.
• ⚡ Optimized the training pipeline for CPU-only environments through backbone freezing, warm-start initialization, and efficient multitask learning.
• 🚀 Deployed an interactive inference demo on Hugging Face Spaces with automated checkpoint management.

Stack: Python · PyTorch · ResNet-50 · OpenCV · Computer Vision · Multi-Task Learning · Hugging Face Spaces · Docker

Cross-lingual question-answering system using multilingual sentence embeddings + FAISS for semantic retrieval. Extractive grounded answering across English, Hindi, Kannada, Malayalam. CPU-efficient pipeline optimized for low-resource environments.

Stack: Python · Sentence-Transformers · FAISS · Multilingual NLP

LLM-driven multi-agent system with Supervisor, Search, Chat, and Worker agents for intent routing, retrieval, and API execution. RAG pipeline with Pinecone for context-aware responses. Conversational memory enables stateful actions like search and booking.

Stack: FastAPI · LangChain · Gemini API · Pinecone · RAG · REST APIs

• 🤖 Reinforcement Learning for LLM-based agents — reward design, OpenEnv APIs, multi-agent orchestration
• 🔍 RAG systems at scale — retrieval evaluation, hybrid search, citation grounding
• 🏗 Production AI deployment — Docker, FastAPI, HuggingFace Spaces, observability
• 📚 Preparing for opportunities in ML / AI Engineering roles

Open to ML / AI Engineering opportunities · Always learning