Private inference at the network boundary.
- Project №
- AIPE-0024-A
- Client
- Financial · Tier 1
- Date
- 2025-09
- Class
- CONFIDENTIAL
- Span
- 14 weeks
- Team
- L. Aguirre · F. Mendez
§ 01 — Brief
A regulated trading desk required model serving inside a private VPC with no egress. The system was specified, built, and handed over in fourteen weeks. Operation is the client’s responsibility.
The engagement was scoped against a single load-bearing constraint. Every subsequent decision is traceable to it. Where alternative paths existed, the alternatives are recorded; the chosen path is signed.
The studio retains audit rights for one quarter post-handover. Filed under AIPE-0024-A · Annex II.
§ 02 — Architecture
Client VPC (isolated) → Inference plane (vLLM · single tenant) → Model registry (read-only · signed)
§ 03 — Results
| Metric | Value |
|---|---|
| p95 latency | 84.0 ms |
| Cost per 1k | $0.0021 |
| Coverage | 99.94% |
Measurements taken against a representative quarter of production traffic. Full traces filed.