Notes from production AI.Not marketing content.
What we learn shipping AI onto client hardware: self-hosting economics, model choices, and security.
Gitleaks now ships security patches only, TruffleHog calls provider APIs by default, and Betterleaks v1.10.1 validates nothing unless asked.

Everything we have
written down.
Every post comes out of work we did: a client system, an internal build, or a benchmark we ran. Filter by topic above, or read straight through.

TruffleHog vs Gitleaks vs Betterleaks: Which to Run in 2026
Gitleaks now ships security patches only, TruffleHog calls provider APIs by default, and Betterleaks v1.10.1 validates nothing unless asked.
NOV 17, 2026
Muse Spark Open VRAM Requirements: GPUs, KV Cache, License
Muse Spark Open NVFP4 weights are 533.0 GB, not 700. KV costs 64 KiB per token plus 368 MiB per sequence. Which 8-GPU nodes fit, and the license steps.
NOV 16, 2026
Hermes Agent vs OpenClaw: Which Can a Company Run? (2026)
OpenClaw has 722 published security advisories; Hermes Agent has none but 39 third-party CVEs. Both say they are single-operator tools. Which one to run.
NOV 13, 2026
DeepSeek Harness Security Review: Egress, Sandbox, Plugins
dsh 0.2.2-rc.2 attaches your session log to every official DeepSeek request by default, and its sandbox never restricts network. Egress table and lockdown.
NOV 12, 2026
Copilot Oversharing: Which Controls Fix Access in 2026
Restricted Content Discovery changes no permissions and can take over a week on big sites. Which Copilot oversharing controls hide files and which close access.
NOV 11, 2026
DAgger Imitation Learning: When Corrections Beat More Demos
RaC beat prior robot results with 10x less collection time. When DAgger imitation learning beats more demos, and what LeRobot v0.7.0's dagger mode records.
NOV 10, 2026
HyDE RAG: When Hypothetical Documents Help or Hurt Retrieval
HyDE RAG took Contriever from 44.5 to 61.3 nDCG@10, yet cut a fine-tuned encoder with 11B and 52B generators. When to gate it and how to A/B test it.
NOV 09, 2026
OpenAI Privacy Filter vLLM: Self-Host It and Decode Spans
OpenAI Privacy Filter on vLLM returns a [T, 33] score matrix, not spans. Pin v0.33.0, restore the Viterbi decoder on the client, and calibrate it on your logs.
NOV 06, 2026
Doc to LoRA: How a Hypernetwork Writes LoRA Weights (2026)
A 309M-parameter hypernetwork turns a document into a rank-8 LoRA on every MLP down projection in one pass. Shapes, objective, costs, and where D2L breaks.
NOV 05, 2026