Leila Haddad
@leilaApplied research → production
Structured outputs, schema validation, and the gap between a demo and a contract.
on Architecture notes: autoscaling GPU inference without melting the bill · 5d ago
Did you measure warm connections or cold starts? Our numbers moved 40% depending on that.
on Tutorial: hybrid search with pgvector and reciprocal rank fusion · 6d ago
What does your alerting look like for this? We can't distinguish 'slow' from 'stuck'.
on Lessons learned from 11,000 agent runs that went sideways · 8d ago
Our p99 was never the model, it was our own JSON serialisation. Took a month to find.
on Tutorial: quantise a model without wrecking its reasoning · 16d ago
Do you version prompts with the code, or separately? We're still arguing about it. (37)
on Lessons learned maintaining an inference engine with 900 open issues · 16d ago
The 'one shared schema' point is doing a lot of quiet work here. (57)
on Tutorial: ship an MCP server your model can actually use · 16d ago
The counter-intuitive part for us was that more retries made availability worse. (3)
on Lessons learned maintaining an inference engine with 900 open issues · 16d ago
Curious how you cost this per tenant — token counts only, or wall-clock too? (13)
on Tutorial: a golden-set eval suite that blocks bad prompt merges · 23d ago
Not sure the added complexity pays off below ~50 QPS, but above that I agree completely. (76)
on Tutorial: a golden-set eval suite that blocks bad prompt merges · 23d ago
How long did it take before the cache hit rate stabilised? Ours needed a full week. (11)