← Steppe Integrations

Articles & Field Notes

Practitioner notes on AI-assisted development, architecture, and integration. Each piece traces back to a specific working artifact: a commit, a load test, a working demo. Receipts over rhetoric.

  1. Field Note

    Four Impossible Things Fell in Two Days. The Fifth Was Real.

    “Unsupported,” “unusable,” “blocked,” and “brutally slow” — four inherited labels dismantled by measurement in forty-eight hours on one all-Intel workstation, including a 40× runtime peel that put a mixture-of-experts model on a 13-TOPS NPU everyone said couldn’t run it. Then the part that makes it trustworthy: the two walls that held, closed by predeclared gates instead of fatigue. Signed predictions, same-day retractions, and the discipline of a lab run with AI agents — theory first, then the case, then the raw numbers.

  2. Benchmarks

    I Upgraded Everything Around the Bottleneck. It Did Not Move.

    A local-AI multitasking ceiling of exactly 8 survived a complete machine rebuild — new CPU, 4× the RAM, new OS, year-newer driver — because it was never hardware. One line of llama.cpp source, found by a twenty-minute discriminating experiment, proven guilty by dose-response across seven name-brand models, and moved: the 9th parallel request gets 6.7× faster on the same Intel Arc GPU. With the patch, the raw data, and a debugging story about a correlation wearing a causation costume.

  3. Benchmarks

    What Consumer Hardware Actually Does

    Twenty-nine indexed local-model labs plus the August rebuild and Qwen 3.8 campaigns on hardware anyone could buy: 12 GB NVIDIA cards and a pair of Intel Arc Pro B70s. The rebuild falsified the simple "buy RAM" answer, exposed an eight-request Vulkan kernel wall, and separated load placement, concurrency and context depth into three different constraints. Includes the 120B that fit, the 30B MoE that became the faster service, local training, corrected retractions and the server that silently truncates an over-limit prompt.

  4. Republication

    Solving the Job Shop Problem

    A 2019 article on job shop scheduling with Google OR-Tools, written from inside a 24/7 manufacturing operation — republished verbatim with a 2026 coda. Same three constraints, same CP-SAT solver, different shop: the machines are now GPUs, the jobs are AI agents, and setup time (loading a model into VRAM) is the constraint that decides everything. Receipts included: 250 jobs, 3 model loads, zero missed deadlines.

  5. Tutorial

    Build Your Own HEARTH

    A local MCP gateway for Claude Code on a Windows machine with an NVIDIA GPU: every tool call logged to a queryable ledger, a local model on tap for the grunt work, and a test gate at every step. ~90 minutes from nothing to receipts. Includes the traps that cost us real hours.

  6. Architecture

    Functional Enterprise Architecture

    The enterprise architecture that actually gets used is smaller and humbler than a binder — a handful of document formats in tools you already know, plus the one command that makes them a habit. The formats that kept earning their place, and why the notes that help you remember turn out to help an agent operate the system cold.

  7. Methodology

    Flowing Water and Stuck Agents

    A debugging methodology for when AI tells you it can't be done. Three patterns — persona-shifted critics, structured pushback, predictive self-critique — from one hard night of multi-card GPU debugging on consumer Windows. Four open-weights frontier labs running on the same box by morning.

  8. Methodology

    Find Every Reason to Dismiss Me

    I commissioned an adversarial AI review of my own code — instructed to find every defensible reason to dismiss it. It accused me of not writing the code. Why that was the most useful thing it could do — and why I published the review, unedited.

  9. Methodology

    The Mech Suit Methodology

    From copy-paste to multi-agent orchestration. Six phases of an AI-paired development workflow, written from lived practice. The AI does not replace your judgment. It amplifies it.

  10. Demo

    threat-trace · Agentic Defense in Your Browser

    A working multi-agent log-analysis pipeline. Three streams under attack, three discrete agents, one cross-stream layer that catches the false-positive trap a single-prompt scanner walks into. No install required.

  11. Field Note

    The bug the trace overlay caught

    A spread-order bug surfaced in two seconds because the in-product trace overlay made the contract violation visible. The case for logs over tests when an agent builds in unfamiliar tech.

  12. Field Note

    The Tight Loop

    Field test for the Mech Suit Methodology. A combat-log parser, a Godot test game, and a hosted multiplayer network became a working 3D replay viewer in two slices over two days. With the receipts.

  13. Build Pattern

    Five Layers for AI-Assisted Build Sessions

    Four parallel AI-builder streams shipped clean in one night. The pattern that made it possible: context tiers, bootstrap protocol, collaboration loop, automation boundary, human-in-the-loop.

  14. Stack

    Dispatch + MAF + OTel: A Multi-Agent Stack

    The complete multi-agent build-and-run stack. Cowork desktop for parallel build sessions, Microsoft Agent Framework for runtime agent orchestration, OpenTelemetry as the observability substrate that spans both layers.

  15. Field Notes · S02

    The Precheck Cut

    Eight field notes from a real repository. The receipts behind the Mech Suit Methodology, with file-path captions and verbatim artifact quotes.