I’m the person behind a new YouTube channel called forwardslashbuilds. I’ve been an in-the-field IT engineer since the late 1900’s, and I’m at a point in my career where I’m making a deliberate pivot: I’m going all-in on AI for practical IT operations.
Not “AI hype.” Not “prompting for vibes.” I mean the kind of AI that helps the working sysadmin, cloud engineer, or MSP consultant ship safer changes, faster, and without turning production into a casino.
Why I’m entering (and why this isn’t a pity story)
I’m not trying to “stay relevant” in the shallow sense. I’m trying to stay useful.
In IT, you either keep learning or you quietly become the person who says, “We’ve always done it this way,” while the world moves on. I’m older, and attempting to become wiser at all turns. Thusly, I have no interest in stagnation. The next edge of the blade is here, and I want to be on it. Openly, in public, with receipts. Bold enough to take the steps, show them, and with the mistakes included.
That’s what forwardslashbuilds is going to be: build logs, experiments, failures, and repeatable wins. Not brand deals. Not guru content. Just the work. I don’t know it all. But have always been crafty and savvy as a jack-of-all-trades in my vertical.
The project: a “Verified PowerShell LLM” for Microsoft cloud + endpoint ops
I’m building an open, modular toolchain I’m calling forwardslashOPS with one clear goal:
Generate PowerShell and operational runbooks that are not just “plausible,” but verifiably correct for the scenario. Using natural language to describe the scenario and goals to execute PS and VBS scripting that works right the first time. As such, I have devoured over 1.5M PS and VBS articles from formal MS documentation to start RAG production proactively.
Think WinUtil, but for the modern Microsoft stack. Azure, Entra ID, Intune, Exchange Online, Microsoft Graph, and the real-world workflows people actually run (device onboarding, compliance baselines, Conditional Access guardrails, RBAC, mail flow, identity hygiene, break-glass procedures, etc.).
But the differentiator is verification.
What “verified” means (in concrete terms)
Most LLM tooling fails sysadmins in predictable ways: confident hallucinations, deprecated cmdlets, missing permissions, wrong module versions, wrong parameter sets, and scripts that “look right” until they touch reality.
My approach is to treat output like software, not fortune cookies:
- Retrieval from authoritative sources:
The assistant will ground responses in official/public documentation and known-good patterns (not private data, not client content, not scraped secrets). - Constrained generation and guardrails:
The model doesn’t get to freestyle. It has to target specific modules, specific scopes, and explicit intent (read-only audit, what-if/dry-run, apply change). - Automated checks before it ever reaches a human
-PSScriptAnalyzer linting
-Pester tests for expected behavior
- WhatIf to then confirm safe defaults where appropriate
-Version checks for modules (Graph SDK, ExchangeOnlineManagement, etc.)
-Permission/scope preflight guidance (least privilege)
- A lab-first validation harness
I maintain a home lab specifically so we can validate actions in a controlled environment before anyone tries them in production.
This is exactly the sort of thing that, done right, becomes a force-multiplier for everyone trying to modernize without burning weekends.
What I’ve already been doing (so this isn’t just an idea)
I’m already building a local AI workflow and a structured knowledge base for IT references. This includes but is not limited in scope at this time to include Windows, AZIdentity, InTune/Endpoint management, and on/offbaord automation patterns. This is because I want this to run locally and privately as much as possible.
I’m also documenting this journey publicly on forwardslashbuilds because I want the learning to be reusable: the setup, the pitfalls, the prompts that fail, the evaluation methods, the “don’t do this in prod” warnings, and of course showing how to make the best workflows that actually work.
What the Radeon AI Pro R9700 enables that I cannot do today (Current sys specs Ry9 9950X3D, 96GB DDR5 6000 [2x48 Crucial Pro 5600 OC to 6k], Ample Gen 5 storage on X870E Hero BTF board, all Noctua thermal management.)
The Radeon AI Pro R9700’s big unlock is VRAM headroom with its 32GB class and pro-focused compute in a form factor meant for real workloads. This is exactly the kind of environment you want for local inference, evaluation, and iteration in one place.
If I win one of the R9700s, it materially changes what I can ship:
Run larger local models at higher quality settings (or run multiple smaller specialist models concurrently) without constant loading/unloading and context thrash.
Create a practical evaluation loop: generate script < validate < test < repair < re-test, at a speed that keeps momentum.
Fine-tuning/adapter training (LoRA-class work) on IT-specific patterns and in a way that’s feasible in a home lab, not just a cloud budget.
Build a multi-model router: one model for intent classification, one for retrieval summarisation, one for code generation, and one for critiquing without turning inference into an exhaustive and burdensome overnight job.
In short: the R9700 turns this from “cool project” into a toolchain I can iterate fast enough to make releases. Ultimately going from individual scripting generation to an LLM that is rich and supportive of all levels of IT from green technical service team members to seasoned and hardened consultants and architects.
Deliverables and timeline (so you can hold me accountable)
If I win, I’ll publish progress in a way the community can verify:
Within 14 days
Repo goes public: Will build Git
Containerized baseline: local inference with retrieval and a minimal CLI
First “verified script” pipeline: generate + lint + basic tests
Within 30 days
Module packs: Entra/Graph + Intune baseline workflows
A starter catalog of “safe automation recipes” (read-only audits + WhatIf-first changes)
Within 60–90 days
End-to-end “Verified PowerShell Assistant” alpha release
Video series: build log, model benchmarks, project failures, ultimate fixes, what we learned
Community-driven backlog: issues, feature requests , and PRs
Brand integrity and boundaries
Two explicit commitments:
- No client names, no client data, no private environments.
This will be built from public references and my own lab artifacts. - No “AI as an excuse” content.
Every script that claims to do something will come with the verification artifacts that prove it: tests, module requirements, and reproducible steps.
In closing:
Level1Techs has always been a place where “do the work” matters more than “say the thing.” That’s why this giveaway hit me: it’s not about collecting hardware; it’s about building something worth showing.
If I win an R9700, it will be used to build a truly open, verifiable AI workflow for IT engineers, and I’ll publish the build so other people can replicate it, improve it, and benefit from it.
Thanks for reading. If you want to follow along:
yt @forwardslashbuilds
