I'm Ashok, a backend engineer from Hyderabad , building AI systems that are measurably correct rather than plausibly correct.
Durable job pipelines, hybrid retrieval, structured LLM routing, and keeping cost per request from quietly eating the margin.

🔭 Currently building NotSudo, an agent that tells you whether a CVE actually reaches your code  ·  🌱 Writing Go from scratch to understand things properly

Project	Stars	Forks	Issues	Pull requests
multea Terminal multi-agent orchestrator for Claude Code: run agents across repos with a unified TUI and DAG task scheduling.
nano-vector-db Embedded vector DB in pure Go, zero dependencies. HNSW from the paper: 0.99 recall@10 at 5,300 QPS on 100k vectors, WAL crash recovery.
notsudo Agent that reasons over real call sites to decide if a CVE reaches your repo, then opens a remediation PR with cited evidence.

🛡️ Adversarial input is an architecture problem, not a filtering problem. In NotSudo, the nodes that read attacker-controlled advisory text are capability-gated and cannot call a single side-effecting tool. The node that opens the PR never sees raw advisory prose. You cannot prompt-inject a node that has nothing to invoke.

💸 Cheap models do most of the work. Two-tier routing kills ~95% of advisories with a small model and saves the frontier model for the one question that actually needs reasoning. Correctness stays; the bill does not.