Applied AI infrastructure, agent systems and reverse engineering.
I build the layer under AI products: model routing, agents that delegate, memory they share, and data pulled out of systems that were never meant to be read. Lately, a product of my own.
Now shipping · my product
Setiap rit sampai tepat waktu.
Every trip arrives on time.
A trip tracker for Indonesian trucking offices. Dispatchers follow every truck on one screen, drivers report through a WhatsApp link with nothing to install, and every finished trip grades the vendor on that route.
Late and at-risk trips rise to the top.
Drivers update from one link. No app, no login.
Multi-drop trips, each stop with its own deadline.
Every trip grades the vendor on that route.
Selected work
J5 EV DashboardReverse engineering2026
CarLinko's cloud sends my car's whole state as one packed blob. I mapped it by driving and watching which bytes moved, then built a self-hosted dashboard on top.
Agent-to-Agent OfficeAgent systems2026
A pixel office where agents hand real work to each other: files, shell, web search. A plain-text tag becomes a real JSON-RPC 2.0 call under Google's A2A protocol.
RouteBenchLLM infrastructure2026
Runs the code each model writes against real tests in three languages, makes it repair a broken repo turn by turn, then exports which model should get which task.
routebench.routing.v1
{
"primary_model": "model-a",
"fallback_models": ["model-b"],
"category_rules": [ ... ]
}memory-coreAgent memory2026
A long-term memory on my home server that Claude Code, Codex and Aider all read and write over MCP. Postgres with pgvector, a local CPU embedder, keyword and meaning search fused.
claude-code asks why does the server keep freezing?
hit · 148 ms · NVMe drops off the PCIe bus; reboot=pci brings it back
9router #1144Open source2026
GPT-5.5 and DeepSeek streamed Read limits like 25999999999999999. Claude Code rejected the call and retried, burning tokens. I buffered the streamed tool arguments and clamped the limit.
- "limit": 25999999999999999+ "limit": 2000
Streamed tool arguments are buffered until complete, then Read's limit is clamped.
How I work
Every project starts from a gap I hit myself: a car app hiding its data, agents forgetting yesterday, trucks nobody could see.
I work through the architecture with AI as a research partner, and write down the decisions before any code exists.
Claude Code writes much of the code. I set the shape, review every change and keep the system coherent.
Nothing ships on the model's word. Receipts, real sessions, real drivers: the result has to match how the world behaves.
Contact