Ten years of backend and platform work. Since 2025 I have built the deterministic layer that lets AI agents do real engineering inside a real company: a compiled graph of the whole system, a governed delivery pipeline, and the evals and telemetry that keep both honest. Below: 6 tools you can install right now, and three essays on how they work.
The work
Every job on this page was a bet that the boring, structural work would matter more than the flashy work. It kept paying. Scroll: the stage on the right follows the story.
Joined a logistics-AI startup as an intern and stayed through its 0→1, from the first customers to a platform other teams built on. Event-driven microservices on GCP and Kubernetes, built while the product was still finding its shape.
Senior engineer on an access-control and smart-buildings platform, where the software ends at a door. I drove R&D on the in-house IoT gateway (AWS IoT Core, MQTT, over-the-air updates), integrated elevator destination dispatch and third-party readers, and built the I/O rule engine: if-this-then-that automation for doors, readers and alarms. Around it I shipped the public APIs and webhook ecosystem partners build on, rate-limited with Redis, the video integrations, and a global dashboard for security operations centres at large multinationals.
Led the migration of 12+ live services from Heroku to AWS Fargate, with Terraform owning the infrastructure and GitHub Actions owning every deploy.
Lead engineer for an AI platform across a fleet of more than a hundred services. Every AI coding tool fails the same way inside a real company: not at writing code, but at knowing the organisation. Which service consumes which event. What breaks three repos away when a column is renamed. Which rules are non-negotiable.
So I stopped asking models to guess and compiled it: a graph of the whole system that answers structural questions in milliseconds and re-checks itself against live source, a spec-driven pipeline agents operate inside rather than around, and evals and telemetry that keep both honest. The engineering team uses it every day.
Tools I ship
5 of them MIT, one free for personal use. Zero dependencies wherever that was possible. Each one exists because I needed it on a Tuesday.
Skills and agents for AI coding assistants, installed into Claude Code, Codex, Cursor, OpenCode, Copilot, Gemini CLI and more from one central clone. Symlinked, so update is one git pull and every harness sees it. Agents transpile to each harness's native format. Includes visual-verify: the agent renders what it built and looks at it before claiming it works.
See where your AI tokens actually go. Reads the logs your tools already write and turns them into honest analytics across Claude Code, Codex, OpenCode, Cursor, Cline and Hermes: 10 adapters, 145 tests, CI on three operating systems. Dashboard, native menu-bar app and CLI. An unpriced model shows up as null with a fix, never as a silent $0. Nothing leaves your machine.
The calendar where your AI agents show up for work. Book a recurring job for Claude Code, Codex, OpenCode or Hermes; it runs unattended inside an OS-sandboxed git worktree that never touches main and cannot read your SSH keys, then files a report you can actually read: what it did, what it skipped and why, what it cost. 13 agent profiles ship with it. Hard caps on dollars, turns and wall-clock.
A video player for what your agent actually did. Turns a session transcript into one self-contained HTML file you can scrub, step through and hand to a teammate. No server, no dependencies, and nothing leaves your machine until you share the file.
Route each task to the cheapest Claude tier that can do it well, through model-pinned subagents. Volume to Haiku, building to Sonnet, review and hard problems to Opus, judgment to the top tier. Only write lanes touch code, and every diff is verified by a different model than the one that wrote it. A task that fails twice escalates with its failure log, never a blind third retry.
An orchestrator and an executor split across two vendors. Claude plans, decomposes and judges; Codex writes the code once the spec is frozen and makes no design calls of its own. Bounded work packets, three executor tiers, an escalation path, and the accepting verdict never comes from the model that wrote the diff.
How I think
Wherever a model could be asked, a lookup is tried first. Models are for judgment. Retrieval, blast radius and contracts are facts, and facts should be looked up, not generated.
The first objection to any knowledge layer is "it will rot." So the layer checks itself: structural answers are re-verified against live source before anyone acts on them.
Published research puts the ceiling for reliable agent tool selection at roughly thirty tools. Fewer, higher-leverage capabilities beat a long menu, and every new one has to justify its slot.
A reviewer that agrees with everything is decoration. Review is structured as disagreement: independent read-only lanes, then a lane whose only job is to refute a passing verdict.
Impact claims stay unpublished until they are measured. Telemetry ships structure, never content. "I could not look" is never reported as "nothing found."
The most useful architecture record is often the thing that was evaluated and declined, with the reasoning written down where the next person will find it.
Writing
Longer form, one idea each. Written for people who build agent infrastructure and want the mechanism, not the pitch.
Seven substrates compiled into one ontology graph, and the drift check that makes a stale answer report itself instead of being trusted.
Read →Why a long tool menu makes agents worse, how the consolidation was done, and what the structured error envelope buys you.
Read →Background agents on a schedule, draft PRs only, a merge capability that is physically absent, and why honest reports were the hard part.
Read →Now · September 2026
Right now I am shipping the tools above and working on the unglamorous part of autonomous agents: making their reports honest, so that "I could not look" never comes back as "nothing found." The next thing I want to publish is a head-to-head benchmark of agents with and without a compiled graph of the codebase, with real error bars.
I have spent ten years on backend and platform work: a founding-team seat at a logistics startup through its 0→1, then access control and smart buildings, where I gave the hardware one gateway and one public API and led the move from Heroku to AWS, and since 2025 have led the AI platform work described above.
Hire me
Agent infrastructure, evals and deterministic retrieval; cloud platforms and the migrations that keep them alive; hardware that has to answer to software. Or a team that needs someone who ships the whole loop and measures it honestly.
vmoksh.shah179@gmail.com · Ahmedabad, India (UTC+5:30)