eric@builds:~$ ./ventures --next
The language layer for IT operations
The console is becoming an API. The API is becoming a prompt. This venture is the piece in between: an LLM that turns an admin's plain-English intent into validated, staged, reversible fleet changes — and explains every one of them back.
Endpoint management has a translation problem. The people responsible for thousands of devices think in intent — "enforce this baseline on the pilot ring," "why did enrollment fail on these laptops" — but the tooling only speaks blades, Graph calls, and log files.
The language layer closes that gap. You say what you want. The model drafts it as policy-as-code, a validator checks it against your guardrails, changes stage to a pilot ring first, and everything lands with a plain-English explanation and a one-step rollback.
The model never touches the fleet directly. It proposes and explains; only the deterministic pipeline executes and enforces. That division is the entire design.
▸Intent → policy-as-code
"Block legacy auth for contractors, pilot ring first" becomes a validated, versioned configuration — with anything ambiguous flagged before a single device is touched.
▸Fleet triage in plain English
Enrollment failures, IME logs, compliance drift — the model reads the logs against a known failure taxonomy and answers the only question that matters: what broke, and what's the fix?
▸Change explainability
Every applied change carries a human-readable rationale: what changed, why, what ring it hit, and how to roll it back. The audit log becomes something a CISO can actually read.
▸Guardrails as first-class objects
Blast-radius limits, staging rings, change windows, and rollback plans are defined up front — the boundaries automation must operate inside, not suggestions it may consider.
▸NQ Terminal — live futures dashboard
A trading console for Nasdaq-100 futures with zero AI at runtime: live chart with TradingView-style zoom, pan, timeframes, and range bars; support/resistance ranked by how often each level actually held today; a 24-strategy analyzer that charges real trading costs; and a confluence engine that only draws a reversal signal when independent evidence stacks — and explains every one. Python standard library and a single HTML file, end to end.
▸why it fits here
Same design rule as the language layer: judgment encoded, execution deterministic. Every signal carries its reasoning, every strategy shows a net-of-costs record, and when nothing qualifies the chart honestly shows nothing. See the footer for the standard futures-trading disclaimer — this is an analysis tool, not advice.
The larger system behind it — a self-managing desk with AI agents on earned authority — has its own running build log: the journey →
the operator layer is already dissolving
Most endpoint work is deterministic, documented, and patterned — exactly what LLMs automate first. The question isn't whether the click-ops layer gets automated; it's who builds the guardrails it runs inside. I'd rather it be someone who has carried a pager for a 150,000-device fleet.
models crossed the reliability threshold — for drafting
Current models are good enough to translate intent into configuration and to summarize state faithfully — if they're boxed in by validation and staging. They are not good enough to hold the blast radius. The architecture assumes both facts permanently.
this is my home field
Twenty-five years of Intune, SCCM, Jamf, and PowerShell means the failure taxonomy, the packaging traps, and the compliance edge cases are already in my head. The venture is that judgment, encoded. The running commentary lives in the field notes.
early & unfinished — on purpose
This page describes an early-stage concept in active development, not a product offering. If you run a fleet and want to pressure-test the idea — or you're building something adjacent — I want to hear from you.