architecture · 2026-05-05
Region routing enforced in code, not policy.
How the routing layer guarantees data residency at the TCP socket level — and why the policy PDF isn't enough.
eng · 8 min read
// writing
Thinking on inference infrastructure.
Architecture decisions, research notes, and operational lessons from running frontier open models at production scale.
infrastructure · 2026-05-12
Why we chose Blackwell over Hopper for 405B-class inference.
The B200 SXM5 is not a generational upgrade from H100 — it is a different substrate for a different workload. At 405B parameters in BF16, the HBM3e bandwidth changes what 'good' looks like in ways that matter to us. Here is the full analysis that justified the infrastructure bet.
legal · 2026-04-28
A plain-language walk through our GDPR DPA.
What each clause means, what we are committing to, and what your DPO should read first.
legal · 6 min read
fine-tuning · 2026-04-18
LoRA fine-tuning cost anatomy.
Where the GPU hours go during a $5,000 LoRA run — GPU utilization, checkpoint saves, eval, and idle.
eng · 10 min read
security · 2026-04-10
What SOC 2 Type I readiness actually involves.
The controls, the evidence, and the six months of infrastructure changes that got us to readiness assessment.
security · 12 min read
api · 2026-03-29
OpenAI compatibility: what we support and what we don't.
A complete mapping of the Chat Completions API — the 95% that works today, the 5% we won't build.
eng · 7 min read
product · 2026-03-18
Why we host fewer models.
Six models at launch vs. 200 in a catalog. The operational reality of serving 405B-class at production latency.
product · 5 min read
infrastructure · 2026-03-08
vLLM on B200 SXM5: first 90 days.
What changed from H100 to B200, what broke, what we patched upstream, and what we're still tuning.
eng · 15 min read
legal · 2026-02-24
India DPDP compliance for AI workloads.
What the Digital Personal Data Protection Act means for enterprises running inference on Indian data.
legal · 9 min read
fine-tuning · 2026-02-10
DeepSeek-V3 fine-tuning on LoRA: benchmark report.
Baseline vs. 5k-sample LoRA adapter on three domain tasks. What moved, what didn't, and at what cost.
research · 11 min read