I turn agent failures into traces, invariants, and regression tests.
AI systems engineering across protocols, evals, memory, and workflow reliability.
Bengaluru 路 open to India, remote, and relocation
Email 路
LinkedIn
01 / External selection. Won the AMD Developer Hackathon 2026 fine-tuning track with BrainConnect-ASD, a site-adversarial GNN evaluated leave-one-site-out: 529 held-out subjects across four sites, ROC AUC 0.7872.
02 / Upstream fixes. Reproduced a live Vertex AI Memory Bank metadata-loss bug and shipped the typed conversion into Google ADK main. For DataHub, replaced destructive incident updates with field-level patch semantics in PR #18826: 17 files, +1,621/-380, with 15 inline review comments. Followed it with idempotent incident creation.
03 / Protocol correctness. Seven merged contributions across AgenTrust's cMCP and conformance suite, including a session-independent execution-correlation specification, nested argument validation, and protocol-correct JSON-RPC handling.
Which memories must disappear when a source becomes untrusted? A provenance and derivation graph for agent memory with selective descendant revocation, signed tool bindings, nonce-based replay resistance, drift detection, and failure injection. 409 passing tests in the current verification suite.
Why did the same API workflow take a different path in two environments? A five-crate Rust engine that records typed execution graphs and diffs the transition, routing, retry, join, and failure-policy decisions that caused the divergence. 249 passing tests.
Which deployment caused an LLM cost or latency regression? A single Go binary that joins model telemetry to deployment SHAs, fetches implicated diffs, and applies deterministic scoring before model-generated explanation.
My ARC-AGI-2 experiments were preregistered and task-group cross-validated. All six learned-verifier variants lost to the frozen baseline, so I stopped the direction instead of post-hoc retuning it.
AI systems engineering at an early-stage AI company, especially agent reliability, evals, protocol correctness, MCP/tooling, and failure-aware workflow infrastructure.