天気
An anime weather forecast that talks
A Unity WebGL chatbot with a lip-synced anime presenter, Indonesian text-to-speech, and a 48-worker batch pipeline for spreadsheets.
Ask Tenki-Chan about the weather in plain Indonesian. An LLM classifies the intent under a strict JSON-only schema, WeatherAPI resolves the location by name or coordinates, and ElevenLabs speaks the answer back while uLipSync drives the character's mouth.
The part nobody expects is the batch mode: drop in a CSV or XLSX of place names and it fans out across a 48-worker pool with retry, backoff and jitter, then hands back a spreadsheet — all inside WebGL, with its own file-picker and download bridge.
It ships as a single WebGL build behind a Terraform-managed Azure load balancer, with Prometheus and Grafana watching, GitHub Actions building the image, and a cache in front of the assets.
where the effort went
the numbers
6000.2.3f1, WebGL only
retry + backoff + jitter
built with
delivery record
I wanted weather data to feel conversational and character-led while still supporting the less glamorous job of checking many locations at once.
Indonesian conversational weather flow with strict structured LLM output
Weather lookup by place name or coordinates
ElevenLabs speech with Unity character lip sync
CSV and XLSX batch processing with forty-eight workers
WebGL file selection and generated spreadsheet download bridge
Terraform-managed Azure delivery with monitoring and CI
The constraints mattered as much as the finished interface.
Named machines, as they appear in the diagram. Yes, they are all named after something.
Drawn as it was built. Open it full size — the labels are the interesting part.
after shipping
Character and voice can make a utility approachable, but the structured data path still has to be deterministic.
WebGL integrations often depend on a narrow browser bridge that should remain smaller than the application logic around it.