AI Developer Tools · Tech Partnerships — Google Cloud
A Java developer who moved into partnerships — and never stopped shipping. Ex-Cloudflare, where I ran deep technical partnerships: product and OEM integrations across security and the developer platform. Today I build the AI developer-tools tech-partnerships practice at Google Cloud, and write open-source agent & eval tooling on the side.
MS Computer Science, Illinois Tech · BTech Computer Engineering · INSEAD MBA · CFA
How I work
I don't hand things off at the hard part. End-to-end ownership and accountability — not the boxes the role defines, but everything the outcome actually needs to get over the finish line.
Fast-moving and half-defined is where I'm most useful. I've worked in engineering, product, partnerships, business, strategy, and special projects — enough angles on the same problem to pattern-match through one that hasn't settled yet.
Building
Vendor-neutral receipts for model routing — records which candidates were known, why one won, and whether the evidence actually supports changing the policy.
A playground built to use Kimi K3 rather than benchmark it — and a scoring bug worth keeping: the strength meter read 100/100 on every plan until it was rebuilt.
Turns an AI coding agent's real production traces into measured, guarded fixes — a fix counts only when replaying the exact failing command goes red to green.
A trust substrate for self-learning agents — makes an agent's own self-modifications verifiable, gated, and reversible.
Keeps an agent's declared constraints alive across context compaction — extract, then re-inject. Cut constraint violations to zero.
Portable evidence receipts for research — independent, reproducible, offline groundedness checks on every cited claim.
An open trust manifest for AI-agent skills — signed attestations of authorship, capabilities, and trust tier.
Synthetic-data-first optical music recognition for notation VLMs can't read — Western staff, Jianpu, and Indian Sargam.
TypeScript agent framework — persistent memory + multi-agent coordination.
An open, vendor-neutral format for AI-agent trajectories — capture from any harness.
Runtime layer that verifies a tool call hit its real intent — not just a 200. Caught 100% of silent failures on τ²-bench, zero false positives.
Small, composable specs for how agents exchange state — memport (memory) · handover (task handoff) · provenant (provenance) · worklease (fleet coordination) · capgrant (scoped capability grants).
Classifies each agent failure — harness bug, reward hack, or training gap — across 5,000 real coding-agent runs vs. ground truth.
On-device AI in the browser — WebGPU + MediaPipe, no server round-trip.
Semantic search over 500+ arXiv papers — Gemini + pgvector on Cloud Run.
All repos on GitHub →Writing
Full-stack AI observability wired across Vertex AI, GKE and the model layer.
Measuring the real productivity lift developers get from AI coding assistants.
Zero-switching-cost model training across any compute, via object storage.
Query a distributed data lake without egress lock-in.
How the Workers/Pages third-party integration surface came together.
Jamstack build integrations for Sanity, Contentful, Ghost and more.
Push-based crawl signaling to cut wasted bot traffic across the web.
All 13 Cloudflare posts →Activity
Recently shipped
46 public repos and counting · github.com/abhid1234