I build production-grade AI systems: the stuff around the model that makes it reliable (data → retrieval → inference → evaluation → ops).
Right now I’m focused on LLM infrastructure + agent reliability:
- reducing “looks done but isn’t” outcomes
- enforcing verification and safe tool usage
- keeping latency and cost predictable
An OpenCode plugin that adds discipline + verification to agentic coding workflows.
- Repo: https://github.com/pkgprateek/setu-opencode
- Docs: https://deepwiki.com/pkgprateek/setu-opencode
What it’s trying to do (in plain language):
- force “context-first” behavior (read before write)
- keep work structured (research → plan → implement)
- make verification non-optional (don’t ship until checks pass)
If you’ve ever watched an agent confidently break a repo, this is my attempt at a practical fix.
- LLM systems: retrieval, routing, evaluation loops, guardrails, monitoring
- Backend and distributed systems: APIs, queues, reliability, performance
- Performance work: latency, throughput, GPU/CPU efficiency, cost control
- Developer tooling: workflows that make engineering calmer instead of chaotic
I default to a few principles:
- Make failure modes explicit (what can go wrong, how we detect it, how we recover)
- Prefer boring, testable infrastructure around “intelligence”
- Log and trace everything worth debugging
- Keep the feedback loop tight (small diffs, quick verification, repeat)
- Issues/PRs on Setu are welcome.
- If you’re shipping an AI product and want a second set of eyes on reliability/cost/perf, I’m open to conversations.
Contact
- Email: hello@prateekgoel.com
- LinkedIn: https://linkedin.com/in/prateekkgoel
- X(Twitter): https://x.com/prateekkgoel