Software Engineer focused on AI systems — agent tooling (MCP/FastMCP), schema-driven backends (Pydantic), and LLM evaluation/fine-tuning. I build schema-driven AI agent tooling and LLM training/evaluation pipelines, especially at the intersection of MCP/FastMCP, backend systems, and blockchain data/RPC.
- 🧱 Building: MCP tool servers + validated schema registries (Pydantic + pytest) for reliable agent workflows
- ⚡ LLM Work: Fine-tuning + evaluation (GRPO/RLHF-style reward shaping, vLLM inference pipelines)
- 🔎 Interests: Agent reliability (validation, retries, tracing), eval harnesses, data quality & contracts
- 🌱 Exploring: LangGraph and LangChain for multi-agent orchestration
- 💬 Ask me about: AI Agents, MCP, LLM fine-tuning, Blockchain & NFTs
- Developed a real-time cryptocurrency monitoring agent using Pydantic-AI and a local Llama 3.2 model, capable of fetching, analyzing, and validating market data from CoinGecko.
- Integrated Supabase for automated storage and updating of top 50 cryptocurrency prices, ensuring accurate and consistent data syncing in realtime.
- Implemented a Model Context Protocol (MCP) server to enable Claude Desktop to perform CRUD operations on the crypto database, using FastMCP and secure environment configuration.
- Designed robust error handling, environment-driven configurations, and optimized SQL schema with indexing for scalable and reliable market monitoring.
- Utilized the Qwen-3-1.7B language model with Group Relative Policy Optimization (GRPO) and Chain-of-Thought (CoT) prompting on a physics benchmark dataset.
- This approach improved the model's reasoning and alignment with human-like problem-solving.
- The integration of GRPO and CoT led to a 12% increase in model accuracy on the physics benchmark.
- Designed a system that integrates Apache Solr's indexed Wikipedia data with AI for efficient knowledge retrieval and summarization.
- Developed a web crawler to scrape and preprocess 55,000 data points and ETL them into JSON format.
- Built a React-based website with DialoGPT for querying, hosted on the Google Cloud Platform.
Languages: Python, Go, SQL, JavaScript ML/LLM: PyTorch, Transformers, HuggingFace, TensorFlow, GRPO/RLHF-style training Backend: FastAPI, Flask, Pydantic, REST/JSON-RPC, GraphQL Infra/Data: Docker, Kubernetes, GCP, MongoDB, Kafka, Spark, Hadoop
-
SIITR: A Semantic Infused Intelligent Approach for Tag Recommendation
- A tag recommendation system using similarity measures and SVM, achieving 94.48% accuracy.
- Presented at ANTIC 2021 and published in Springer. : https://link.springer.com/chapter/10.1007/978-3-030-96040-7_31
-
ISBRNM: Integrative Approach for Semantically Driven Blog Recommendation Using Novel Measures
- A blog recommendation system using NLP, GRUs, and optimization techniques, with 95.85% accuracy.
- Presented at ICDTA 2022 and published in Springer. : https://link.springer.com/chapter/10.1007/978-3-031-02447-4_2
