Supernomics
A 45-second cold start was costing users and stalling enterprise deals. We cut it to 9 seconds and hardened the AI/ML pipeline into a security posture buyers could diligence.
Supernomics runs an AI workforce product, and a 45-second cold start was a tax on every user and a drag on every enterprise deal. The AI/ML pipeline also needed a security posture solid enough to survive a buyer's diligence. Both had to be fixed without slowing the roadmap.
How we took it on.
- 01
Profiled the pipeline to isolate cold-start cost from steady-state latency.
- 02
Re-platformed from Cloud Run to GKE for warm-pool control and predictable scaling.
- 03
Hardened the AI/ML pipeline with secrets management, network segmentation, and observability via Langfuse.
What shipped.
GKE-based serving with warm pools and autoscaling tuned to traffic shape.
Secured ML pipeline with end-to-end tracing and evaluation.
Observability and alerting wired to SLOs from day one.
The outcome, in plain terms.
Cold-start latency fell from 45s to 9s, an 80% reduction users feel on every request.
A security posture buyers could diligence, aligned to SOC 2.
Predictable scaling under real traffic, operated with the pager held.
- Google Cloud
- GKE
- Cloud Run
- Langfuse
- gRPC
- Prometheus
- SOC 2
- Secure SDLC
- SRE
Senior engineers who owned the problem end to end, with the latency numbers to prove it.