Jeff Sperandeo.
Forward Deployed AI Engineer
I take AI from a demo to a system that runs inside a real business.
Live and public
CarVector: a vehicle-data API and MCP server I built and run.
Specs, federal recalls and diagnostic codes, with a native MCP server for AI agents. Usage numbers are read from the live service when the page loads.
- 12,344+
- vehicles
- 25,849+
- federal recalls
- 1,284+
- OBD-II codes
Demo: sample requests against the public API
Don’t take my word for it. Send the request — this hits the live API I built.
Discipline as a feature
Aetherframe: legal-document AI that cites everything and advises nothing.
It reads a case file page by page, organizes every document, builds the timeline, and cites every fact to its source. A refusal layer, enforced in code rather than in a prompt, blocks predictions of who wins and any legal strategy. That rule is the product: it is the only safe way to put AI next to real legal work.
Judgment, not just code
The backup custodian
On its first day, the backup system I built found a production server with zero backups, and fixed it before anything was lost. Judgment built into the system, not bolted on after.
Running in production
A dozen live products, one platform, one engineer.
A service business’s whole AI stack
Veteran Vehicle Repair, since 2023: leads, booking, and live phone AI that transcribes every call, scores it, and turns it into a follow-up with no staff data entry.
A multi-tenant platform
One process serves a dozen products, each with its own isolated database. A new business is one config file.
An autonomous content engine
One model writes the claim, a second checks it claim by claim, and a gate decides what publishes. It runs a 24/7 channel, scheduled by the machine.
An AI video pipeline
Script, voice, render, mix and publish, end to end. Any AI-generated person is labeled, consented synthetic media, in line with the FTC rule of October 21, 2024.
A research copilot
Runs its own code, keeps memory across sessions, and reasons over a private knowledge base. A system, not a chat box.
nine gates · one keystroke each
Most people ask AI for an answer and paste it. I run every change through a gauntlet — including a pass where I assume my own code is wrong and try to break it — and nothing ships that didn’t clear all nine gates.
That’s the job of a forward-deployed engineer: get a real system into a real business fast, without the speed becoming a liability. This pipeline is how both are true at once.
Most people live in one layer and call a vendor for the rest. I work in all six — so when something breaks at 3am, I know exactly which floor to go to.
their field · my lab
In their field
On-site with the people who’ll use it. The goal isn’t to start building — it’s to find the gap between the problem they described and the problem they actually have. This is the stage a contractor skips.
In their field
Translate the ambiguity into a spec and a definition of “working” — the eval criteria, the success metric, the acceptance bar. Locked with them before any code, so “done” isn’t a moving target.
In my lab
Pull back into my environment and build production-grade — the integration, the AI wiring, the guardrails. Fast, but written for the engineer who maintains it next (me, in three weeks).
In my lab
Stress-test against the messy edge cases real data throws — not the happy path. The eval harness lives here: does it hold up when the input is garbage, adversarial, or just weird? If it breaks here, it never leaves the lab.
In their field
Deploy into their environment with a limited blast radius — one team, one workflow, reversible. The field test, deliberately small so failure is cheap.
In their field
Watch it run on reality, catch where it breaks, iterate. Then document and template it so it scales without me babysitting — which frees me to walk to the next problem. The field teaches what the lab couldn’t.
Embed→Scope→Build→Eval→Pilot→Harden — then the field feeds back into Build. A loop, not a line.
Guardrails
Fast, without the speed becoming a liability.
- Idempotent jobs — re-running never double-charges or double-sends.
- Fail-closed gates — when authorization is uncertain, the answer is no.
- Spend caps — every paid model call is metered and announced before it runs.
- Typed errors, zero silent failures — when it breaks, it says exactly what broke.
From police work to AI
I was always the one who could not leave a system alone.
2008
My first real site went up.
I was always the quiet computer guy who could not leave a system alone.
2009 – 2020
Law enforcement.
I became the one who worked the cyber cases and computer forensics. I was commended for recovering a victim’s stolen property through preservation requests and subpoenaed records. When the department moved off paper tickets, I was its liaison: the state trained me, and I had the BADGER TraCS digital citation platform live in officers’ hands in 48 hours against a one-month plan.
2020 – now
Video production, then fractional CMO / CTO.
Video production from 2020–21 on, which became the AI media pipeline. Since 2023, marketing and technology for Veteran Vehicle Repair as its fractional CMO / CTO.
Now
Forward deployed AI engineering.
I saw where AI was going and taught myself every layer, much of it at 4 AM. Self-taught, no degree: judge the running products.
The role
What I’m looking for
- A Forward Deployed Engineer or AI deployment role, embedded with the people who will use what I build.
- Work that goes from an unclear problem to a system running in production, and stays running.
- A team that wants AI shipped with guardrails: cited sources, spend limits, fail-closed checks.
Why the role exists
Why the role exists
7%
of organizations say they have fully scaled AI across the enterprise.
McKinsey, The State of AI 2025 (survey June 25 – July 29, 2025; n = 1,993)
800%
growth in Forward Deployed Engineer job postings, January – September 2025.
Not a rebranded solutions engineer: this role ships production code and owns it long-term. The role started at Palantir, where they called them “Deltas.” For reference: OpenAI posted a San Francisco FDE range of $162K–$280K plus equity.