Mine · 2025
Mathmatika
I built a maths learning workspace with stable streaming equations, generated animations, and a voice mentor.
Streaming a maths explanation token by token caused every incomplete formula to re-render and flicker.
The learning surface needed to remain readable while the model was still responding. Generated animations also needed a recovery path when model-written code failed, and document retrieval needed evidence before adding more infrastructure.
The decisions that made it work
Batch the stream at 80 characters
Formula-aware batching instead of token by token rendering. I spent a week on the naive version first, which is how I know the flicker is unrecoverable for trust.
Generate the animation, then heal it
Claude writes the animation code, a subprocess renders it, stderr is parsed, and failures retry against a template fallback. Cached by hash so a video renders once.
I benchmarked my way out of a vector database
BM25Plus over the uploaded documents scored well enough that the vector database never got installed. Less infrastructure is more product, and I nearly skipped the benchmark that proved it.
0
vector databases
1
LLM provider end to end
2
self-healing retries max