Agentic orchestration
Designing multi-step agent systems: planning, delegation, and control flow across tools and sub-agents. Most of the difficulty is in knowing when an agent should stop, not what it should call next.
Shivajith Mutteal / GenAI Engineer
Building and evaluating language model systems that hold up outside the demo.
I work on GenAI systems — agentic orchestration, MCPs, guardrails, and RAG middleware — and build smaller things on the side because I want to. This site is the lab: side projects, experiments, and notes along the way. Most of it built to be read, forked, or clicked through, not just described.
Selected work
04Guardrails for agentic systems
Bounding what an autonomous system is permitted to do, and detecting when it drifts outside those bounds: validation at the edges, permission boundaries, and containing failures before they compound.
Tools and MCP integrations
The interface layer agents act through: tool definitions, Model Context Protocol servers, and the schema design that determines whether a model reaches for the right thing at the right moment.
Skills
Packaging procedural knowledge so agents apply a consistent approach across tasks rather than re-deriving one each time.
IISc
01Capstone work from the DLFA programme at the Centre for Continuing Education, IISc Bangalore. Finished and submitted — written up here in more detail than the report had room for, including what we got wrong.
Computation offloading with deep reinforcement learning
Capstone, team of four. Three phones, one edge server, one agent per device on its own local view — every millisecond each device decides how much power to spend computing locally versus offloading over a fading uplink. We built the MEC simulator, then dropped two learners into it: DDPG, which can set the power dial anywhere, and DQN, which has to pick from five notches. Writing it up two years later I found three defects in the code, fixed them, and re-ran it over multiple seeds — which turned the original blanket result into a more honest, workload-dependent one.
Live apps
02Eval Playground
Run the same prompt across models side by side and compare output, latency, and cost per call in one view.
Glass-box RAG
A retrieval-augmented-generation demo that makes the retrieval step visible — keyword, semantic, and hybrid side by side, with similarity scores and the near-misses just below the cutoff. Runs entirely in your browser.
Deployment
01Guides and infra-as-code for standing up your own free-tier infrastructure — the fiddly parts solved once, so you don't have to.
oracle-vm-always-free
How to get a 4 OCPU / 24 GB Ampere ARM VM on Oracle Cloud's Always Free tier and keep it running forever, at $0/mo — Terraform config plus a retry script for the common out-of-host-capacity error that trips most people up.
Lab
03Smaller things I built because I wanted to, not because a job needed them.
side project
Chess opening trainer
Spaced-repetition drills for the openings I keep blundering, built after losing the same line four times in a row.
side project
Trail route generator
Give it a distance and a starting point, get back a runnable loop instead of an out-and-back.
weekend build
Recipe remix bot
A just-for-fun tool that riffs on whatever's actually in the fridge instead of sending you shopping.
Writing
03What breaks first when you scale a RAG pipelineJun 2026
Notes on evaluating LLM evalsApr 2026
A small classifier beat our 70B model — here's whyFeb 2026