I'm a Jakarta-based developer who ships end-to-end products alone — from spec documents and data pipelines to frontend, deployment, and the test suite that gates it all. My work spans three worlds that rarely overlap: civic technology, news & media intelligence, and algorithmic trading.
What ties them together is one architectural rule I apply to every project — whether it's a digital human, a news pipeline, or a trading system. It's easier to show than to state.
That is the entire program. One line of arithmetic, applied to itself, forever. No dataset. No model. No training run.
There is no image file behind this. Each frame re-derives itself from the rule, on demand, at whatever depth you ask for.
Same arithmetic, same result — on any machine, in any year, at any zoom. That reproducibility is what makes a system testable, and what makes its output defensible.
Pure, typed functions with no network and no clock inside them. Every calculation — burst scores, viseme timings, position sizing, geo resolution — is reproducible, versioned, and covered by a CI-gated test suite before anything ships.
Language models narrate results, summarize, and select from pre-validated candidates. They never produce ground truth. That's what keeps an AI product honest in domains where people treat the numbers as sacred — because sometimes they are.
Hyper-realistic digital humans of Indonesian figures — powered by a curated local knowledge base — brought to life on the web as talking avatars and on stage as holograms. Flagship: Gus Dur AI, endorsed by NU & PKB.
A news aggregation and trending engine covering Indonesia's major markets — Jabodetabek, Surabaya, Bandung, Medan, Makassar — with regional attribution and a deterministic burst scorer producing bounded 0–100 outputs.
Dashboard anything with AI. Point it at your data and get a live, queryable dashboard — with the deterministic layer computing every metric and the model handling only the natural-language layer on top.
Multi-timeframe market analysis frameworks, execution tooling with discipline validators and risk-based position sizing, and options pricing engines — each shipped with its own reference implementation and test suite.
A private legal AI appliance for Indonesian law firms, notaries, and corporate legal teams. It runs entirely on local hardware, so sensitive case files never leave the office — research, contract review, and drafting without shipping documents to a public cloud.
Zero-capex industrial rooftop solar for Indonesian factories and warehouses. We fund, build, and operate the array; the site owner pays only for the electricity it produces under a Solar PPA — no capital outlay, no maintenance burden.
Projects start as locked specification documents — design, compliance, build order — with numbered invariants. Code that violates an invariant doesn't merge.
Anything that computes a number is deterministic and replayable. Versioned parameters trigger replay when they change. This is what makes bugs findable and results defensible.
Indonesian regulation — UU PDP for personal data, UU ITE for published claims, UU Hak Cipta for sourced content — is woven into the spec from day one, not patched on before launch. Attributive framing can be a lint rule.
I build AI-heavy products, and precisely because of that, the model never produces ground truth. It summarizes, translates, and selects from candidates the deterministic layer already validated.
I take on end-to-end builds — especially products for the Indonesian market where data rules, published claims, and trust actually matter. If you need something engineered, not just generated, let's talk.