ANDROID / VEHICLE OPS Live on Play Store
IntelliAuto
Vehicle management. AutoMind AI, ML Kit OCR, offline encrypted storage.
KotlinFirebaseOffline-firstML Kit
SYS.01 Systems engineering · local AI · product
Linux infra, local AI, shipped apps.
Enterprise Linux operations, LLM inference on Intel Arc Pro B70, and production Android apps — documented with the trade-offs left in.
Enterprise Linux
Lifecycle, automation, hardening, HA and the operational work that keeps platforms predictable.
Select a node to inspect the system
01
Local AI Lab
One GPU, measured to the token.
Intel Arc Pro B70 32GB. Each card is one engine, one model, one cell — not a blended leaderboard.
Nemotron 3.5 Lightning 186.6 t/s C1 client post-first · p2048/g128 · n=5 · 150 W vLLM XPU · GPTQ-INT4 · DFlash n=7 Not isolated prefill. Cold input 7160 is prompt/TTFT. Muse Glimmer-30B 26.8 t/s C1 engine decode · p512/g128 · 128K · 230 W llama.cpp SYCL · Q4_K_XL · DFlash n2 Public path. vLLM overlay is experimental and slower. Qwen3.6-35B-A3B 70 t/s C1 decode · MoE 35B-A3B llama.cpp SYCL · Q4_K_XL Prefill 1659 t/s at the published pp4096 cell. KV cache 0.003 KL Near-lossless vs FP16 KV · ~50% VRAM llama.cpp · q8_0 K + q4_1 V q4_0 K is catastrophic (KLD ~5.5). Do not use it.
02
Production Apps
Useful software, not AI wrappers.
Two privacy-minded Android products built around local-first value and resilient cloud boundaries.
ANDROID / FINANCE OPS Live on Play Store
IntelliFlow
Finance tracker. Offline-first, cloud sync, budget projections, spending summaries.
KotlinFirebasePrivacy-firstForecasting
03 Educational Series
The LLM Infrastructure Handbook
An engineer's guide to Local AI. Demystifying Transformer architectures, KV cache quantization, modern attention and MoE routing.
04
19 B70 index Local LLM Recipes
vLLM XPU · 3B-active LatentMoE Nemotron
186.6 t/s C1 with DFlash, not native MTP
Graphs plus a BF16 DFlash draft. Native MTP still accepts 0%. Speed card is 16K n=5; capacity completed 119,904 tokens. Not 128K.
llama.cpp SYCL · dense multimodal Muse Glimmer
26.8 t/s at 128K — vLLM is not the public path
DFlash n2 wins the GGUF sweep. The 29.2 number is 64K. INT4 vLLM exists only as a local overlay.
llama.cpp SYCL · dense 27B + MoE 35B Qwen
The living recipe, then the 27B MTP card
Qwen numbers stay on their own posts. Do not mix them with Nemotron client-TTFT or DFlash cells.
KV · power · context ceilings Methods
Why q8_0 K + q4_1 V, and what 32 GB actually holds
Methodology posts. These are not model speed cards.
Brand logos are trademarks of their respective owners; shown for identification.
05
Rest of posts
More field notes
01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60