Skip to content
View brej-29's full-sized avatar

Block or report brej-29

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
brej-29/README.md
Brejesh Balakrishnan. AI Engineer, agentic AI and evaluation.

AI that knows when it's wrong. Agent governance on Amazon Bedrock and AgentCore. Guard models, RAG that declines, evals first. Building with Claude Code, Codex and OmniRoute.

Portfolio LinkedIn Hugging Face Email Profile views

About me

An animated terminal types out: whoami, Brejesh Balakrishnan, AI Engineer at Cloud202, governance and safety for agentic AI on Amazon Bedrock and AgentCore. Then about.yml, the agent-govbench summary and the projects he is working on now.

A rotating sphere of glowing tokens with packets orbiting it

AI Engineer at Cloud202 since Sep 2026, working on governance and safety for agentic AI systems on AWS. Before that, about two years as a full-stack engineer at Accenture, modernizing enterprise Java systems for a Fortune 500 financial services client.

  • πŸ›‘οΈ Agent safety: guard models that check a tool call before it runs, red-team suites, policy-aware agents
  • πŸ§ͺ Evaluation first: the test plan and thresholds are written before the model, with confidence intervals on every headline number
  • πŸ”Ž RAG that declines: cited answers from source text, and a plain "not covered" instead of a guess
  • 🧠 Fine-tuning: LoRA / QLoRA on Gemma, Mistral and small routers that run in a browser
  • πŸš€ Shipping: FastAPI, MCP servers, LangGraph, and free-tier deploys on Hugging Face, GitHub Pages and Cloudflare

AI workbench

Animated diagram: Claude Code, Codex and the Claude app send requests through the OmniRoute AI gateway, routed by Nemotron, out to Claude, GPT, Gemini, Gemma 4, Llama 3.3, Mistral and Nemotron. A scrolling strip shows where the work ships: Amazon Bedrock, AgentCore, Hugging Face, Kaggle, Groq, Google Cloud, GitHub Pages, Cloudflare Workers, Streamlit and MCP servers.

Claude Code Codex OmniRoute NVIDIA Nemotron Gemini Gemma 4 Llama Mistral Amazon Bedrock Hugging Face LangGraph Model Context Protocol

Measured work

Animation: proposed tool calls stream toward a guard. Harmful calls turn red and stop, safe calls turn green and execute.

Agent GovBench. Attack success falls from 51.5% with no guard to 4.7% with an LLM guard, on live agent runs.

Code Β· Dataset

Toolcall Guard. Blocks 0.62 of attacks at a 0.08 false-positive rate. 8-bit matches full precision, 4-bit loses recall.

Live playground Β· Code

Compliance Workbench. 677 tests passing, 5 MCP tools. RAG over EU and US regulation that declines instead of guessing.

Live app Β· Code

Mozhi. 0.941 action macro-F1 on 2,805 test requests. A Tamil, Malayalam and English voice assistant with a 2 ms trained router.

Try it in the browser Β· Code

Feasible Day. 96 of 96 feasible day plans, versus 2 of 576 for an LLM alone. Weak spot: 59% on the routing re-check in 15 added cities.

Live app Β· Code

Clearcut. 85.3% fraud recall when only 0.5% of payments can be reviewed. 64 of 75 frauds caught in 284 reviews.

Live workbench Β· Code

🧩 More live projects (click to expand)
Project What it does Open
🦺 SafetyOps Copilot Turns a safety observation into an explainable assessment, an owned corrective action and a shift handover. Read-only MCP tools. Live
πŸ“‰ Customer Churn MLOps Quality-gated XGBoost champion, FastAPI contracts, SHAP-grounded LLM explanations with faithfulness checks, CI gate. API
πŸ“¦ DemandCast Browser-local replenishment: daily sales CSVs in, an explainable order out, tested on 184 held-out weekly decisions. Live
πŸ”­ Kindred One title you loved in, its kin across anime, manga, novels, movies and series out. NDCG@10 0.217 vs 0.142 baseline. Live
β™ŸοΈ Gambit Realms Chess where every capture is a duel, eight measured-strength opponents, and a coach that warns before a blunder. Play
🎨 Toonforge Private photo-to-cartoon and print studio. Local computer vision, optional LLM direction, MCP tools. Live
🌀️ Outside Checks every hour of a planned outing against your weather preferences and explains the closest alternative. Live
πŸ›Ά Riverbeat Rhythm game: row a Kerala snake boat through the backwaters at sunset, one beat at a time. Play
πŸ“– Three Little Stories Pop-up picture books for young children: a monsoon boat, a morning kolam, a cloud who cannot sleep. Read

Earlier work: Analytics Copilot (Mistral-7B QLoRA text-to-SQL) Β· Feast fraud feature store Β· FinAI real-time market analytics Β· SageMaker inventory classifier Β· Streamlit apps

Tech stack

Python, PyTorch, TensorFlow, scikit-learn, OpenCV, FastAPI, Flask, Django, Java, Spring, TypeScript, JavaScript
AWS, Google Cloud, Docker, Kubernetes, GitHub Actions, Cloudflare, Vercel, Linux, PostgreSQL, Redis, Kafka, MongoDB
React, Next.js, Vite, Angular, Tailwind, HTML, CSS, MySQL, SQLite, Git, Postman, Jenkins

LangChain RAG PEFT and LoRA MLflow XGBoost Optuna Pinecone Groq Streamlit Kaggle Power BI

GitHub stats

GitHub stats Most used languages

Contribution streak

Profile details and contribution history

Repositories per language Most committed languages Productive time of day, IST

GitHub trophies

Contribution activity graph

Terminal-style GitHub stats, regenerated daily

Contribution arcade

3D contribution calendar

Pac-Man eating the contribution graph Snake eating the contribution graph

πŸŽ“ Background and credentials
When What
Sep 2026 to now AI Engineer, Cloud202. Governance and safety of agentic AI on AWS (Bedrock, AgentCore)
~1 yr 9 mo Advanced Application Engineering Analyst, Accenture. Enterprise modernization in Java / Spring Boot + Angular for a Fortune 500 financial services client
2024 B.E. Electronics and Communication, College of Engineering Guindy, Anna University. CGPA 8.54. Final-year project: a wearable navigation aid for visually impaired users

Certifications: AWS Machine Learning Engineer Nanodegree Β· Oracle Cloud Infrastructure Data Science Professional Β· Advanced Data Science & AI (Logicmojo) Β· McKinsey.org Forward Program

Contact

Email LinkedIn Portfolio Hugging Face

Evidence first. Then the claim.

Pinned Loading

  1. rag-agent-workbench rag-agent-workbench Public

    Production-grade RAG chatbot with a FastAPI + LangGraph backend (Pinecone vector search + Groq LLM + Tavily web fallback) and a Streamlit chat UI, secured via API key and observable in LangSmith.

    Python 1

  2. otaku-oracle otaku-oracle Public

    Anime answers checked claim by claim against AniList. Live on free tiers: otaku-oracle.brejesh-bala.workers.dev

    TypeScript 5

  3. kindred kindred Public

    Kindred: free cross-media recommender. Name one anime, manga, manhwa, movie or series you loved, get kin in every medium. Pre-registered benchmark, runs in your browser.

    Python 1

  4. safetyops-copilot safetyops-copilot Public

    Safety observation triage, corrective-action tracking and shift handovers. FastAPI/Postgres workspace, explainable DistilBERT severity, YOLO PPE pipeline, and read-only MCP tools.

    Python 1

  5. agent-govbench agent-govbench Public

    Independent benchmark of guardrails for a tool-using AI agent on Google Cloud: Model Armor, rules and LLM guards on 999 synthetic scenarios

    Python

  6. toolcall-guard toolcall-guard Public

    Distilling a small Gemma 4 guard model for AI agent tool calls: LoRA adapter, sanitized dataset, replay evaluation and honest results against Gemini guards

    Python