Headline

Agentic AI | AI Engineer | React/React Native + AWS, now building LLM agents, RAG & multi-agent systems (LangGraph, MCP) from scratch | Python, FastAPI | LLM evaluation & prompt-injection defense | Open to Work

About

I spent three years building React and React Native apps end to end, including the AWS serverless backends behind them. Then I got curious about what was actually happening inside AI agents — and decided the only way to find out was to build one from scratch. No LangChain. Just a loop, a provider-agnostic client, tools generated from type hints, and every failure mode discovered the hard way. Then I kept going. What I've built recently: docagent — a multi-agent research assistant over ~100 technical PDFs that answers with verifiable citations. The hard part wasn't the agents; it was ingestion. Naive RAG flattens a PDF into text, which destroys tables — and in research papers, the answer is usually in a table. So I parse each document into typed elements and chunk each type differently. A planner splits questions into sub-queries, workers search in parallel, and a critic verifies every claim has a source before the answer ships. miniagent — an LLM agent framework built without an orchestration library: autonomous loop with three independent runaway guards, prompt-injection-hardened tools (I red-teamed my own agent and found a credential leak in my file-read tool), retries, tracing, and measured p50/p95 latency. Later rebuilt in LangGraph specifically so I could articulate what the framework gives you and what it hides. How I work: I measure instead of assuming. When I added hybrid search to my retrieval pipeline, I benchmarked it against a golden set rather than trusting that it helped — and found it's situational, not a free upgrade. When my agent cited real numbers under the wrong dataset, I traced it through five layers down to how LaTeX floats tables. That instinct comes from my other work: adversarial evaluation of frontier models, where I plant false premises and verify every claim against source before crediting it. I'm looking for agentic AI / AI engineering roles where systems have to actually work in production — not just demo well. Bengaluru or remote. Python (FastAPI, asyncio, Pydantic) · LangGraph · RAG & hybrid retrieval · MCP · agent evaluation · React / React Native · AWS · Docker

Location

Mysore

AI stack

anthropic apis

claude code

claude

codex

langgraph

claudecode

Work experience

AI Model Evaluation at Freelance | Self-Employed, Jan 2026 - Present · 11 mos, Bengaluru, Karnataka, India, Frontier-model evaluation & adversarial testing

Software Engineer at Urneeds, Apr 2022 - Jul 2025 · 3 yrs 4 mos, Mysuru, Karnataka, India

Education

PGDIT, Big data, Jain (Deemed-to-be University)

Bachelor's degree, Computer Science, Amrita vishwa vidyapeetham,mysore campus

Public links

github: https://github.com/Yashwanth-Brahma