Second year, BSc Data Science & AI, Leiden · everything here was built next to coursework
SilaWay is a companion for the sıla yolu — the long drive families make from Western Europe back to Türkiye by car. Fifteen tools on a bounded loop, four autonomous systems behind it, one engineer. The interesting part isn't what it knows. It's what it does when it doesn't.
Kapıkule'de sıra ne kadar? Bu akşam mı geçsem, sabaha mı bıraksam?
How bad is the queue at Kapıkule? Cross tonight, or leave it to the morning?
Flip it. Same question, same tools — one of these two answers gets someone stuck at a border with a sleeping kid in the back.
~900downloads
last month
Not one agent — four autonomous systems, each with its own failure mode, cost profile and schedule. They don't share a runtime; they share a contract. Pick one and take it apart.
Voice and text, streaming. Routing, border forecasts, fuel, weather, and place-finding from vague language — “şu benzinlikten sonra sağda bir yer vardı” has to resolve to coordinates.
Every tool answers through the same versioned envelope, so the model never sees a bare number. It sees a number and what that number is worth — and the wording it is allowed to use follows from that. Pick one and read the last line first.
All fifteen tools on the loop. The input schemas and the system prompt stay in the app.
planning the drive
the road ahead
finding somewhere
putting it on the screen
This is the whole design in one rule. “We couldn't measure it” is not the same as “it's fine”, and the distance between those two sentences is a family sitting in a car for four hours.
So the board below has a state most dashboards refuse to have: blank. When a sample ages out, the number doesn't decay gracefully into a guess — it disappears, and the agent loses the right to quote it.
Recreated board, real crossings. Live data is in the app.
A Cloud Run service fans out through Cloud Tasks to the official traffic authorities of four countries along the corridor, then normalises, clusters and trust-scores what comes back before an LLM writes the report.
Its first live snapshot was 45 German roadworks and zero congestion. Correct, and useless to a driver.
Nothing was broken. The pipeline had faithfully ranked by how sure it was, and it was very sure about roadworks, because roadworks are published as tidy structured records months in advance. The fix wasn't a model change — it was calibrating severity across sources and ranking by traveller value instead of confidence.
What the first snapshot handed a driver heading for Kapıkule tonight.
A job mines road-trip videos every 30 minutes: audio → transcript → structured report. A local relevance gate reads the transcript and drops the clip before the report step, because that step is the one that costs money.
Raw media and full transcripts are never kept. What survives the job is a hash and a report — a job that runs 48 times a day shouldn't also be building a private archive of other people's holidays.
A second scheduled job pulls news feeds and uses a small model to decide what is actually about the road, rather than merely about the country.
—
Shared by all four systems
All model traffic goes through one Cloudflare Worker. No API key ever reaches the device, spend is capped per user and globally, and every call is accounted for in D1 — 24 migrations of accounting, because the alternative is finding out at the end of the month.
Serving. Turn 18 of 60 — the loop has room.
Illustrative limits. The real ones live in the Worker, with an admin dashboard on top.
9modules,
9 suites
SilaWay is closed source, so I extracted the unglamorous half: the part that has nothing to do with prompts or models, and everything to do with what happens when a tool returns nonsense, when a stream is cut mid-tool-call, when the user cancels. Every module in this list exists because something went wrong first. Click one.
Team project · May 2026 · Node core API + Python FastAPI matching service
AI companion & matching app — the mobile client, and work on the API behind it
3platforms,
1 codebase
Nobody fills in a form. You talk to Dostie, and the profile falls out of the conversation — a pipeline reads what was said, builds a profile from it and proposes matches. I built the client half of that loop, on one Expo codebase shipping to iOS, Android and web. Watch it happen, then cut the connection.
profile, extracted from the conversation
0% complete
A field only appears once something was said that supports it, and a later message merges with the earlier one rather than overwriting it — a person who codes until 2am and runs at dawn is not a contradiction to be resolved, it is two facts. values stays empty for the whole conversation, because nothing in it was about values. Cut the connection halfway: the messages queue, the chat drops to polling, and the profile keeps filling.
Delft, Netherlands
Turkish (native) · English (B2) · Leiden, class of 2028
2products shipped,
one of them alone
TypeScript
The language I've actually put in front of users. Two products, one of them written alone.
The part I'd rather you heard from me
I've built a client against a FastAPI service and shipped the contract between us. I have not been the person who gets paged when that service falls over. It's the gap I most want to close, and I'd rather take a job where I'm the least comfortable person in the room than one where I already know the answer.
{
"v": 2,
"tool": "assess_candidate",
"status": "partial",
"data": { "lang": "python", "level": "coursework" },
"confidence": 0.35,
"freshness_s": 0,
"note": "never owned a service in this language"
}Internships, part-time engineering work, or a problem you think I'd like
Delft, Netherlands · studying in Leiden
↓I read
everything