About

What I build

At work: production backends across FinTech compliance, document AI, enterprise conversational AI and high-concurrency campaign platforms. Most of that time has gone into a GST compliance and document-AI platform, where I'm a primary contributor to the core domain service and the extraction worker — ledgers with correct tax math, reconciliation with the government portal, return filing, durable notifications, and the boundary between tenants.

On my own time, things one step further out: a pull-request reviewer built on a code graph, a memory architecture where retrieval is a decision rather than a reflex, and co-maintaining Loams at Ostrium Labs.

How the work changed

I started with the parts of a system that decide whether it's correct under load and under failure, and kept running into questions underneath them that don't have answers yet: when computation is worth doing, how much verification a result needs, what a deadline does to the value of an answer.

A scheduler for jobs whose value depends on when they finish turned into Temporal Compute, whose first registered experiment lost to EDF. A question about verified results under a fixed budget turned into FRONTIER. The most useful thing I've learned to do with an experiment is to write down what it should show before it runs, and keep the result on the page when it doesn't.

Now

• ReviewGraph — the deterministic half is built. Next is the evidence verifier, which has to exist before the system is allowed to publish a finding on its own.
• FRONTIER — re-running the allocation comparisons under measured costs, after exp-009 showed most of a verification is fixed overhead.
• Temporal Compute — Generation II is registered and waiting on its execution freeze.

As of 28 October 2026

Where

1. May 2026 — nowSoftware Engineer, jpteam / C1X Inc.Full-time
2. Aug 2025 — Apr 2026Software Engineering Intern, jpteam / C1X Inc.Internship
3. OngoingCo-maintainer, Ostrium LabsLoams · Ferite docs and brand