The first thing you'll see is your idea, half-built, not an invoice.
Most freelancers send you a quote based on a Zoom call. I send you a personalized demo video of your project working, within 48 hours of your brief. You iterate on the demo, not on a contract.
I build a working sketch of your idea before we agree on anything.
After you send a brief, I record a 2–3 minute walkthrough of your project running: a tiny interactive prototype, the architecture I'd use, the parts I'd push back on, and a rough price. You react. I refine. By the time we scope the real build, you've already seen (and felt) the thing.
The five steps from brief to ship
- 01/Brief
You send me a brief
Through the quote form: what you're building, who it's for, why now, and any constraints (budget, timeline, tech). Five minutes to write.
- 02/Demo video
I send a personalized demo video within 48 hours
Loom-style walkthrough with a working prototype of your idea, my proposed approach, and the parts I'd challenge, plus a rough price. If we're not a fit, I say so here.
- 03/Feedback
You send feedback
What landed, what didn't, what you want different. A voice memo or a few bullets is enough. The point is to react to something concrete, not negotiate over a doc.
- 04/Iterate
I refine the demo
Usually 1–2 rounds. Once the demo direction is right, we agree on scope, milestones, and price for the real build. No surprises at signing.
- 05/Ship
I ship, with you in the loop
Bi-weekly progress demos so you see the build evolve. Source code in your GitHub from day one. Documentation and handoff included.
- Founders with a real idea and a working budget who want to see it built, not described
- AI / LLM products: agents, RAG, MCP, FHIR, the things I do every day
- Bilingual (EN / FR-CA) projects, especially Canadian SMBs and Quebec-based teams
- Projects between $5K and $80K where I can commit 2–8 weeks of focus
- Static brochure / Wordpress sites without dev work, local agencies do this faster and cheaper
- Sub-$5K work, the demo-first process doesn't pencil out at that scale
- Design-only, no implementation, I build, I don't just deliver pixels
- Projects requiring on-site work in a specific city, I'm remote-only
“Can't I just vibe-code this myself?”
Fair question, and my honest answer is: yes, you can, and you should use AI as much as possible. What I sell is the seam. Here are four things AI pair-programming shipped in my own builds this year, every one of which type-checked and every one of which broke in production.
01“The code runs. Why do I still need you?”
Because “it runs” is the lowest bar. On my last hackathon (a maternal-triage MCP server for PromptOpinion), I spent about 20 hours across 15 commits chasing why a dashboard rendered blank in production. AI-authored components used the wrong prop names for the partner platform's widget wire format. Every Approve / Reject button silently failed a Zod schema check.
On my laptop with test data, it looked fine. In front of judges, every button was dead. That is what an AI pair-programming session produces without a human who has read enough production incident reports to know which wire mismatches fail quietly.
02“But the AI is really good at debugging too, right?”
Sentry's own AI (Seer) diagnosed exactly a memory leak in their Python SDK that I was patching upstream. Then it proposed a fix that would silently drop every builtin exception from deduplication. The prose plan and the actual generated patch were two different things.
AI review is a force-multiplier when a human is doing the accepting. It is not a replacement for the human who says “wait, this patch does something different from what it says.”
03“Can AI hallucinate something that actually hurts me?”
Yes, and I have watched it happen on a medical build. Before I wired the maternal-triage system to real NICE clinical guidelines, Claude Sonnet 4.6 confidently made up the severe-hypertension threshold (170/110 instead of the real 160/110 under NICE NG133) that decides whether a pregnant patient goes to labour and delivery.
That single hallucinated number would have shipped as “AI-powered maternal triage” and missed preeclampsia cases in production. The AI does not know it does not know. The same is true of every domain your business runs on: pricing rules, compliance thresholds, refund windows, licensing terms.
04“What about security? The AI wouldn't do something dangerous, would it?”
On that same medical build, the top-level error handler printed the raw Error object to logs. For every axios HTTP failure, that meant the full Authorization: Bearer token and the PHI-adjacent response body ended up in log storage the AI had already claimed was PHI-free in the README. The README said one thing, the code did another.
I caught it, patched it, moved on. A vibe-coded version of the same feature ships to production and quietly becomes an incident when someone reads the logs.
AI produces code that runs. I produce code that survives contact with your customers. That is the seam you're paying for: the judgment that says “this Zod discriminator collides with the prop named type,” “this state transition needs a guard,” “this default value silently overwrites the clinician's input.”
If you can describe it in a paragraph, I can show you a working sketch in 48 hours.
Send me your brief.
Working demo back within 48 hours. If it doesn't land, you owe nothing.