Systems that explain themselves.
guessing is easy. explaining yourself isn’t.My flagship, Brok’s Forge, is an AI Engineering Operating System — it keeps the reasoning behind an AI system, not just what happened to it. How it actually answers a question is the part worth scrolling for.
- Software Engineer @ TCS
- Google Cloud · BigQuery
- 2 systems live in production
SHEET 01 — THE LOOP
runs on every merge122
REST endpoints
499
tests passing
5
architecture layers
now building: closing the CI test-skip gap on the kernel and forge-* modules
fig. 01 — the four pillars. Every screen belongs to exactly one.
- 122REST endpoints shippedBrok’s Forge V2, OpenAPI/Swagger
- 499backend tests passing83 test classes, JUnit 5 + Testcontainers
- 10Maven modulesframework-free kernel + forge-* platform
- 2systems in productionBrok's Forge · Kuriosity
Brok's Forge — The AI Engineering Operating System.
Most AI tooling tells you what happened. Brok’s Forge is built to tell you why your system is the way it is — the decisions behind it, the evidence for them, and the reasoning that connects the two, answered deterministically.
the résumé is the repo →Brok doesn’t guess. It composes.
In plain terms: the answers are provable and repeatable, not AI guesses you still have to go verify yourself.
Four product pillars — every screen belongs to exactly one.
- Build
- Evaluate
- Understand
- Evolve
- 122public REST endpoints
- 499backend tests passing
- 34end-to-end browser tests
- 802Java files — 663 main · 139 test
Things production taught me.
Four real defects, diagnosed to root cause and fixed — the actual chain of reasoning, not the highlight reel.
The pipeline that had never once run
CI/CDLesson A workflow that "shows green" and a workflow that "has ever executed" are different claims. Check both.
The health check that checked nothing
CI/CDLesson Exit code 0 is not evidence anything happened.
502s after every successful deploy
InfrastructureLesson A successful deploy of one container can silently break a caching assumption in a container you never touched.
115 corrupted characters, 23 files
CorrectnessLesson Fix the mechanism, not the symptom — and prove your own guardrail actually guards.
Kuriosity — the other side of the loop.
Brok's Forge evaluates AI systems; Kuriosity is one — a live RAG product I built and operate end to end. Building both is the point: I understand the platform and the workload it serves.
Kuriosity inverts the AI-tutor model: an AI named Kurio plays the student. Upload study material and Kurio asks context-grounded questions, scores your free-text answers against the retrieved source chunks, and tracks topic mastery across sessions.
Custom RAG pipeline
PDF parsing, chunking, Cohere embeddings, ChromaDB similarity search — every response grounded in the uploaded source, not the model’s general knowledge.
LLM evaluation loop
Free-text answers scored against retrieved chunks — accuracy %, missing concepts and revision topics returned as structured JSON.
Cross-session memory
PostgreSQL tracks per-user topic strength, so weak concepts resurface until they hold.
Operated in production
JWT auth with email OTP, rate limiting, and a three-platform deploy (Vercel, Railway, custom DNS/TLS) run under real traffic.
- React
- Node.js
- Express
- PostgreSQL
- ChromaDB
- Cohere
- Groq
Applied computer vision and systems fundamentals — the range behind the platform work.
Violence Detection & Alert System
Real-time CV pipeline that detects violent activity in video streams and triggers automated safety alerts — YOLOv6 + MobileNetV2 over live footage.
- Python
- YOLOv6
- MobileNetV2
Road Safety Surveillance
Detection and tracking of road-safety violations in real time, replacing constant human monitoring with an automated pipeline.
- Python
- OpenCV
- ML
Student Management System
A clean object-oriented CRUD application for student records — early proof of the same instinct: model the domain first.
- Java
- OOP
Not just two demos.
Brok's Forge and Kuriosity are the proof. Everything below is what makes them possible — shipped in the platform, the products, or the day job. No aspirational logos.
Backend & runtime
- Java 21
- Spring Boot
- Spring Security
- PostgreSQL
- Redis
- Flyway
- REST / OpenAPI
- Node.js
AI systems
- LLM APIs (Claude · GPT · Gemini · Groq · OpenRouter)
- RAG pipelines
- embeddings & vector search (Cohere · ChromaDB)
- evaluation harness
- prompt / dataset versioning
- LLM-as-judge metrics
Infrastructure & ops
- Docker
- AWS EC2
- Nginx + TLS
- Prometheus
- Grafana
- structured logging
- health probes
- OpenTelemetry-ready
- CI/CD (GitHub Actions)
Data & cloud
- GCP
- BigQuery
- Looker
- LookML
- SQL optimization
- row-level security
Frontend
- React
- Next.js
- TypeScript
- Tailwind
- Vite
- TanStack Query
Practice
- clean architecture
- modular monolith
- ADRs
- schema-first migrations
- multi-tenant security
- SSRF / IDOR hardening
- documented testing strategy
Principles I refuse to compromise.
I'm an AI engineer who treats reasoning as something worth engineering, not just answers. By day I build governed cloud data platforms at Tata Consultancy Services; nights and weekends belong to the workshop — most recently, Brok's Forge, an AI Engineering Operating System built around one hard rule: no language model in the reasoning path.
Knowledge should be derived, never authored
If a fact isn't a byproduct of real work — a promotion, an evaluation — I don't want it in the system. Anything a human has to remember to write down eventually stops being true.
Absence is not health
An unmeasured thing and a passing thing are different claims. I'd rather a system say 'unknown' and be honest than say 'green' and be guessing.
Evidence beats confidence
A plausible-sounding answer I still have to go verify hasn't saved me any work. If I can't trace a sentence back to the row that produced it, I don't trust the sentence.
Concede before you’re caught
I'd rather publish a comparison page that admits a competitor is more mature than get caught overstating it later. Confidence without a caveat isn't confidence — it's exposure.
The schema is the contract
Databases outlive code. Append-only migrations and immutable versions, so a result means the same thing a year from now as it does today.
Fix the mechanism, not the symptom
When 115 characters corrupted across 23 files, I reversed the exact mis-decode instead of pattern-matching broken strings — then negative-tested my own guardrail to prove it actually guards.
The way I build AI didn’t come from AI.
It came from three worlds, built before this one.
An island with nothing to win — only a place to wander, and cows to pet.
A story about being bullied online, lived from inside the person it happens to.
A room where physics became something you could reach out and hold.
I never changed the way I think.
Only the problems I chose to solve.
From the workshop.
On the bench now
Closing the CI test-skip gap — the kernel and forge-* modules currently install with tests skipped in CI, so their own suites don't run there yet.
2026
Brok's Forge V2 shipped — Brok (zero-LLM reasoning), Root Cause Explorer, AI Git and Forge Graph, on top of the original registry and evaluation engine. 122 REST endpoints, 499 backend tests, 34 end-to-end specs.
2026
Diagnosed and fixed a chain of production incidents to root cause, back to back — the full write-ups are a few sections up, in Field Notes.
2025
Repaired repository-wide UTF-8 corruption and shipped a negative-tested CI guard against it — also in Field Notes.
2025
Kuriosity shipped — a RAG tutor in production across three platforms. Still running.
The work behind the workshop.
Governed cloud data platforms by day; open-source AI infrastructure the rest of the time. Both are production systems with real users.
Brok's Forge — Independent / Open Source
2026 – PresentCreator & Maintainer · AI Engineering Operating System · RemoteCurrent
- Designed and built Brok's Forge V2, an AI Engineering Operating System that records the decisions and evidence behind an AI system and answers engineering questions about it deterministically — no language model anywhere in the reasoning layer.
- Architected a five-layer platform (Forge Kernel → Registry → AI Git → Forge Graph → Engineering Applications) on a deliberately framework-free, append-only event kernel with its own API/core/TCK/PostgreSQL-adapter split.
- Built and hardened a fully unattended production deployment to AWS EC2 (Docker Compose, Nginx, Let’s Encrypt) with automatic image-tag rollback on a failed health gate, across 9 GitHub Actions workflows including CodeQL and dependency review.
- Diagnosed and fixed a chain of live production incidents to root cause — a CI/CD pipeline that had never once executed, a health check silently skipped by a stdin-consumption bug, and Nginx serving 502s by caching a stale upstream IP after every deploy.
- Java 21
- Spring Boot 3.4
- PostgreSQL 16
- Next.js 15
- AWS EC2
Tata Consultancy Services
Dec 2025 – PresentSoftware Engineer · Google Cloud / Looker · ChennaiCurrent
- Design and develop production LookML models, views and explores on Google Cloud Looker, enabling self-serve analytics for business teams.
- Optimized BigQuery SQL, reducing average dashboard load time by ~35%.
- Built 10+ Looker dashboards across Finance, Operations and Marketing.
- Implemented row-level security and access-group policies for multi-tenant data governance.
- GCP
- BigQuery
- Looker
- LookML
NIT Trichy
Sep 2024 – Oct 2024Game Developer Intern · Trichy
- Shipped a full game on online-bullying awareness with Unity and C#, owning design through deployment.
- Unity
- C#
MIC Chennai
Aug 2023VR Developer Intern · Chennai
- Built immersive VR experiences with Unity's XR Interaction Toolkit, improving simulation fidelity through iterative testing.
- Unity XR
- C#
SRM Institute of Science & Technology
2021 – 2025B.Tech, Computer Science Engineering · Trichy
Credentials, quickly.
Google Cloud — Associate Cloud Engineer
Google Cloud
Java (Certified)
HackerRank
JavaScript (Certified)
HackerRank
Python
CodeChef
AR/VR Game Development
CDAC Pune
Build something with me.
Open to AI engineering and platform roles — teams that care whether the system still works six months after the demo. If that's you, the inbox is open.
i read everything. usually same day.