AI-native systems builder · Remote + travel

Jesse LeVasseur

I design, ship, and operate production software end-to-end — by directing AI tooling, fast, and solo.

Four years in public safety, rising to fire captain. In 2026 I taught myself to build with AI and shipped a production estate in months — roughly 15 Cloudflare Workers, 7 databases, self-hosted GPU inference with two fine-tuned models, telephony, and more, across three live platforms. I don't hand-write code from memory; I architect systems, drive the build with AI tooling, and own them in production.

Based Whitewright, TX Work Remote · travel welcome Focus forward-deployed · AI application · founding builder
01

What I build

Architecture & cloud

Serverless / edge systems on Cloudflare — Workers, D1, R2, scheduled jobs, deploy pipelines. API & auth design, including an OAuth 2.1 server built from scratch.

Applied AI

LLM application design, local + cloud model integration, fine-tuning (LoRA), self-hosted serving (vLLM), and document-extraction / RAG-style pipelines.

Data & automation

Web scraping / ETL, structured extraction from messy government and legal documents, database and domain modeling.

Integrations

Payments (Stripe), telephony with a browser softphone (Twilio), email and social capture — wired into real operating workflows.

02

Selected work

Public-Records Analytics Platform

Founder & sole engineer2026 — present

An end-to-end platform that automates open-records (FOIA) requests and turns government documents into structured data — architected and operated solo on Cloudflare's edge stack.

  • Built the full request pipeline (D1, scheduled Workers, an operator console, a statute library, an automated send planner) and scrapers across board-meeting and records systems spanning 1,200+ government bodies.
  • Fine-tuned a Qwen3-14B "PIA-lawyer" LoRA served on a self-hosted GPU, wrapped in defense-in-depth guardrails — arithmetic forced into code, a citation whitelist, a false-proposition denylist, a human-review gate. Measured 96% flag recall, 100% citation grounding.
  • Reads scanned government records with a self-hosted vision model (Qwen2.5-VL) at ~100% title accuracy, replacing commercial OCR; fabrication-proof per-trustee vote extraction from board minutes.
  • Whole-domain email on Cloudflare (DMARC-aligned), a Twilio phone line with browser softphone, and structural chain-of-custody (hash → archive → read-back → re-hash) for evidence integrity.
Cloudflare WorkersD1R2LoRA fine-tuningvLLMvision/OCRPythonTwilioscraping/ETL

Parliamentary Governance Platform

Founder & engineer2026

A governance SaaS that encodes real deliberative procedure — motions, voting, minutes, bylaws — in software.

  • Clause-anchored comments and a redline lifecycle on governing documents; resolutions; a 2/3-threshold ballot + eligibility voting engine; a full minutes lifecycle; capability-based roles; an append-only public audit log.
  • An OAuth 2.1 authorization server built from scratch for an AI (MCP) connector; hand-rolled RFC 8291 Web Push; iCal and client-side PDF export.
  • Reused my records-platform architecture to stand the whole thing up in under two weeks.
single-Worker + SPAD1R2OAuth 2.1 (from scratch)MCPdomain-driven design

National Library-Data Census

Founder & engineer2026

A free, open data census of public-school library catalogs, with an AI review pipeline.

  • A scan engine running concurrent pure-HTTP checks across 6 reverse-engineered catalog (ILS) platforms, each hit cited to its exact catalog URL, behind a misattribution-integrity filter for shared-SaaS hosts.
  • Review pipeline on a local 14B model plus cloud models (~$0.09/record) with a byte-exact verbatim-grounding gate so tags can't be hallucinated.
  • Re-architected analytics onto a cost-optimized local mirror after a $160/month cloud read blowout.
HTTP scan enginelocal + cloud LLMcost optimizationStripedata pipelines
03

Engineering highlights

04

How I work

A pilot who can fly the plane — not rebuild the engine from memory.

AI-native, by design. I architect the system, direct the full build with AI development tooling, then test, debug, and run it in production. I ship working systems fast — I don't compete on writing algorithms from memory on a whiteboard.

Outcomes over pedigree. Everything above is live and operating. I built it solo, from nothing, in months — reusing my own architecture to move even faster the second and third time.

Ownership under pressure. Before software, I ran fire and EMS calls and led disaster-response deployments. I bring the same composure and total ownership to shipping.