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.

Explore field notes

system_map://sergiiob.dev

interactive
01 / 04 PLATFORM
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 running llama.cpp SYCL — every number below is hardware-measured, not estimated.

70 t/s MoE 35B decode @ 512K ctx Q4_K_XL · 1659 t/s prefill 29 t/s Dense 27B + MTP-4 @ 256K +50% speculative gain 0.003 KL q8_0 K + q4_1 V cache ~50% VRAM, near-lossless 512K max context on 32GB MoE reads 3.8× less KV
02

Production Apps

Useful software, not AI wrappers.

Two privacy-minded Android products built around local-first value and resilient cloud boundaries.

ANDROID / VEHICLE OPS Live on Play Store
IntelliAuto app icon

IntelliAuto

Vehicle management. AutoMind AI, ML Kit OCR, offline encrypted storage.

KotlinFirebaseOffline-firstML Kit
Play Store intelliauto.app
ANDROID / FINANCE OPS Live on Play Store
IntelliFlow app icon

IntelliFlow

Finance tracker. Offline-first, cloud sync, budget projections, spending summaries.

KotlinFirebasePrivacy-firstForecasting
Play Store intelliflow.work

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.

Read the Handbook
04

Recent Field Notes

Measured, failed, fixed, documented.

64 View Full Archive
Latest signal

MTP Spec-Token Curve on Intel Arc Pro B70 — 204.6 t/s Decode at N=4 (and Why 97% Acceptance Isn't the Goal)

The MTP draft head is recurrent, so a single-layer MTP checkpoint emits N speculative tokens per step. The spec-N curve (N=1/2/4) is the real story: N=4 wins throughput at 204.6 t/s single-stream decode (+41% vs the community 145 claim) even though per-token acceptance falls with N (92.6% → 84.2% → 72.7% → 68.9% by draft position). N=1 reaches 97.1% acceptance — but costs 30% throughput. The right objective for a single-layer draft head is throughput, not acceptance %. A paired, alternating power A/B also killed the 'boost prefill to 230W' hypothesis: 150W vs 230W prefill is flat at ±0.2%.

b70 local-ai infrastructure
02

Intel Arc Pro B70: vLLM vs llama.cpp — Corrected MoE + Dense Showdown

b70 local-ai infrastructure
03

Phase 1: The vLLM Question on Intel Arc Pro B70 (MXFP4 Native Test)

b70 local-ai infrastructure
04

Intel Arc Pro B70: +140% Long-Context Prefill from a llama.cpp Master Build

b70 local-ai infrastructure
05

Intel Arc Pro B70: The Complete Local LLM Recipe

local-ai infrastructure b70
06

Running Qwen3.6 27B on Intel Arc Pro B70: The Full Recipe

local-ai infrastructure b70