The architecture behind this portfolio — one shared WebGL canvas, MDX content validated at build time, and a deliberate rule that adding anything is one file.
· 2 min read
- next.js
- architecture
- three.js
Backend & AI Engineer · Hyderabad, India
I build backends for AI products — and the evals that keep them honest.
Founding engineer at Metry AI, where I own the backend architecture and REST APIs behind SOJO — a client-management platform for beauty and wellness businesses across Asia. Outside work I build agent systems and the evaluation harnesses that tell me whether they actually work.
- Founding engineer · Metry AI
- Agentic RAG in production
- 38 of 50 on SWE-bench Verified Mini
- 9 ingestion pipelines, one knowledge base
- Judge panel + a gate that blocks its own deploy
- FastAPI · Spring Boot · LangGraph
- 27 claims on this site, each with a receipt
- Hyderabad · IST (UTC+5:30)
Selected work
All workmacOS Troubleshooting Agent
Case studyResolved 2 of 4 multi-turn cases faster than GPT-4o (n=4) — and an eval that's just as honest about where it loses.
Solo — design, build, deploy, evaluate · Python · LangGraph · FastAPI · Qdrant
Coding-Agent Harness
Case study38 of 50 resolved on the clean run — and 14 of the first pass's 24 failures turned out to be the grader, not the agent.
Solo — design, build, benchmark · Python · LangGraph · Docker · Modal
Finance Dashboard API
Case studyThe cleanest Java in the portfolio — a real security layer, split DTOs, and integration tests, not just a CRUD demo.
Solo · Java · Spring Boot · Spring Security · Spring Data JPA
E-Wallet Microservices
Four services, two consumer groups, and honest gaps named up front — no idempotency, no DLQ, not yet a production system.
Course project — solo · Java · Spring Boot · Apache Kafka · Spring Data JPA
Reliable Video Transfer over UDP
416 lines below the framework layer — the only project in the portfolio that isn't built on someone else's abstraction.
Solo · C · POSIX sockets · pthreads
This Site
Adding a project is one MDX file. The 3D layer ships zero bytes to a browser that can't render it — poster instead of canvas, every time.
Solo · Next.js · TypeScript · Tailwind CSS · MDX
I’m a backend engineer in Hyderabad, currently a founding engineer at Metry AI. I like the parts of a system that are unglamorous and load-bearing — and I build agent systems on the side mostly to find out where they break.
More about meWhat I work with
Languages
- Python
- Java
- TypeScript
- JavaScript
- C
- SQL
LLM & agents
- LangGraph
- Agentic RAG
- ReAct tool loops
- Tool calling
- LLM-as-judge evaluation
- SWE-bench benchmarking
- Sandboxed agent execution
- Guardrails / PII redaction
- Provider failover
- Embeddings & vector search
Backend
- FastAPI
- Spring Boot
- Spring Security
- Spring Data JPA
- REST API design
- Microservices
- Kafka
- Event-driven architecture
- JWT / OAuth
Data
- PostgreSQL
- MySQL
- Qdrant
- SQLAlchemy
- Alembic
- JPA projections
Infra & ops
- Docker
- GitHub Actions / CI
- Prometheus
- Grafana
- Loki
- Oracle Cloud
Frontend
- React
- Next.js
- TypeScript
- Tailwind CSS
- MDX
- React Three Fiber
Systems
- POSIX threads
- UDP socket & protocol design
Experience
Mar 2026 — Present
Remote — async with a Canada-based team- Own backend architecture and REST API development for SOJO — — carrying microservices and data models from design through production support.
- Cut by reworking database indexing, caching, and query paths to eliminate redundant PostgreSQL reads.
- Integrated LLM APIs and rebuilt prompt workflows behind structured testing and evaluation, raising response quality and consistency across platform features.
- Instrumented API performance, service health, and production failures in Grafana, shortening the path from a failure firing to a diagnosed root cause.
- Work largely async with a team about ten hours away, across architecture, CI/CD, and reliability decisions.
- Python
- TypeScript
- Node.js
- PostgreSQL
- REST APIs
- Microservices
- Docker
- GitHub Actions
- Grafana
Writing
All postsTwo out-of-phase trig waves multiplied together — the cheapest displacement that doesn't look like it repeats.
· working- glsl
- sketch
About the ✓ marks
Claims on this site carry a receipt. Click one and you get what backs it and a link to the exact file, so you can check me instead of taking my word for it.
There are 27 of them. 18 point at code, committed benchmark data, or something running. The other 9 come from private work and link nowhere — those are marked — instead of ✓, because a claim you can’t check should not look like one you can.
A script enforces it. It fails the build on unbacked superlatives, on a benchmark percentage without its sample size, and on any claim that says it’s checkable but links nowhere. It’s the same idea as the regression gate on my macOS agent, pointed at my own writing instead of a model.
Want to talk about a backend, an agent, or an eval that keeps you honest?
Email is fastest — I read everything.
cheetiharichandra2@gmail.com