Backend engineer with 3+ years of experience building scalable systems and extensively working with production-grade Generative AI applications.

• Strong in Python backend + LLM infrastructure (RAG, agents, MCP, streaming, real-time systems)
• Building complex GenAI systems that handle real constraints (latency, rate limiting, state management)
• Focused on performance, scalability, and production-ready code
• Experience with Claude API, LangChain, LangGraph, vector databases, and MCP frameworks

• Agentic Workflows → multi-step reasoning, tool orchestration, dynamic decision-making
• RAG Systems → semantic search, document intelligence, retrieval optimization
• LLM-powered APIs → Claude integration, streaming responses, cost optimization
• Real-time AI Systems → low-latency decision making, event-driven architectures
• MCP Servers → structured tool definitions, API integration, scalable tool abstractions

Python • Java • SQL

FastAPI • Flask • Django • REST API Design • Async/Await • Event-driven Architecture

• LLM Frameworks: Claude API, LangChain, LangGraph
• Retrieval & Search: RAG pipelines, vector databases, semantic search, hybrid search
• Agent Architecture: ReAct pattern, tool orchestration, state management, reasoning loops
• Infrastructure: Model Context Protocol (MCP), streaming, prompt engineering
• Fine-tuning & Optimization: Prompt optimization, cost reduction, latency tuning

AWS (Lambda, API Gateway, S3, Bedrock, DynamoDB, CloudWatch, ECS Fargate, SQS) • Docker • PostgreSQL with Vector Extensions

PostgreSQL • MongoDB • DynamoDB • Vector Databases (Pinecone, Weaviate concepts)

Problem: Make trading decisions in <30 seconds using real-time market data under API rate limits.

Claude-powered agent for SENSEX/Nifty 50 options using MCP. Combines price action reasoning with real-time data orchestration.

Implementation highlights:

• MCP Server Design: Built structured tool definitions for market snapshots, candle analysis, options pricing, technical indicators
• LLM Integration: Claude reasoning loop for entry/exit decisions using price action analysis (not lagging indicators)
• API Constraint Handling: Solved Angel One rate limiting (50 req/min during live hours) by migrating to Dhan API (25 req/sec, 250/min, 5000/day)
• Real-time Architecture: Async event-driven backend with <500ms response times for decision-making
• State Management: Persistent trading rules and decision history for pattern recognition across trades
• Production Rules: Trading rulebook with bounce trap detection, double-bottom retest patterns, OI-based strike selection

Why this matters:

Shows you can build LLM systems under real constraints: time pressure, API limits, state persistence, and production-grade decision logic.

Repository: https://github.com/SomuTech/realtime-options-agent

Problem: Semantic search and Q&A over large document collections without full document re-reads.

End-to-end RAG system for intelligent document search, answer generation, and document intelligence.

Implementation highlights:

• RAG Pipeline: Document chunking, embedding generation, vector storage, semantic retrieval, LLM generation
• Retrieval Optimization: Hybrid search (semantic + keyword matching) for precision improvement
• Performance Tuning: Chunk size and overlap optimization reduced hallucination rate and improved F1 score
• Backend Scalability: FastAPI service with PostgreSQL for metadata, vector indexing for semantic search
• Production Features: Caching layer, concurrent request handling, error recovery, response streaming

Why this matters:

Shows you understand modern RAG patterns, retrieval engineering, and production LLM application architecture.

Repository: https://github.com/SomuTech/genai-experiments/tree/main/smartdoc_assistant_RAG

Problem: Automate complex research tasks requiring multi-step reasoning, tool usage, and iterative refinement.

Multi-agent system for research automation using agentic workflows and tool orchestration.

Implementation highlights:

• Agentic Loop: Planning → tool execution → reflection → iteration using LangGraph
• Tool Orchestration: Dynamic tool selection and chaining for research workflows
• State Management: Persistent reasoning state, memory management for long-running workflows
• Extensibility: Plugin architecture for adding new tools and agents without code changes
• Error Handling: Graceful failure recovery, fallback strategies, timeout management

Why this matters:

Shows you can build complex agentic systems with robust state management and production-grade error handling.

Repository: https://github.com/SomuTech/genai-experiments/tree/main/deep_research_AGENT

• Agentic AI architectures and ReAct patterns
• LLM infrastructure, scaling, and cost optimization
• Real-time constraint systems with LLM reasoning
• MCP-based tool abstractions and integrations
• RAG systems and retrieval optimization
• Production GenAI application design

• Generative AI Engineer roles (LLM infrastructure, agents, RAG)
• AI Systems Engineering roles (agentic workflows, real-time systems)
• Backend Engineer roles at AI-focused companies

Interested in teams working on hard problems at the intersection of LLMs, systems design, and real-time constraints.