I build working
AI production
systems.
Not decks about them.
Applied AI, mostly across marketing and CRM. I look for where AI changes the economics of a job, prototype fast, then ship something built on real data for a specific context — end to end, through security review, into people's hands. Built by hand, not bought off a shelf. Three of those systems are below.
TorontoAI · Marketing & CRMSolo builds, shipped
Built & shippedAgentic workflowsProduction AIMarketing & CRMConversational analyticsGenerative mediaData infrastructureText-to-SQLPersonalizationPrompt systems
Built & shippedAgentic workflowsProduction AIMarketing & CRMConversational analyticsGenerative mediaData infrastructureText-to-SQLPersonalizationPrompt systems
Focus areas: Built & shipped, Agentic workflows, Production AI, Marketing & CRM, Conversational analytics, Generative media, Data infrastructure, Text-to-SQL, Personalization, Prompt systems.
Selected workThree projects
01Working AI production pipelineLive
AI Content Studio
One product photo in, a multi-format retail campaign out.
Browsable end to end · live generation gated
Every render spends real credits, so generation is behind a passcode. Internal colleagues can reach out directly; recruiters, message me on LinkedIn and I will open it up.
A campaign studio that turns one product photo into stills, bilingual EN/FR promo tiles and video. A packshot generator that produces GS1 planogram angles without a reshoot. An Ad Lab of preset ad recipes — editable, reference-locked, with the sound built in layers the way a studio actually does it. And a prompt builder that teaches the structure rather than handing over a prompt.
Thirteen models from six labs behind two APIs, routed by what each is actually good at: reference-to-video where the packaging must not drift, a cheap draft tier where it does not matter yet. Costed per render before you spend — including the token-billed models, where resolution moves the price more than length does.
Generative AI·Video·Production systems·Next.js
02Consumer AI productMarket ready
Persopot
Selfies in, studio headshots and outfit try-ons out.
Pricing and static try-on demo open — generation needs an account
Two AI products on one trained identity: studio headshots ($29–$79 one-time) and an outfit try-on subscription that composites any Pinterest pin or retail product image onto the user's trained face ($5/mo, 30 credits).
Solo build. Two parallel ML pipelines — FLUX.1 for headshots, FLUX.2 for outfits on fal.ai — share a Gemini 2.5-pro validation gate that catches identity drift before users see it. Next.js, Supabase, Stripe, Cloudflare R2 and Trigger.dev: about 95 API routes and 16 migrations across payments, generation and social features (collaborative outfit boards, “ask a bestie” reviews, wishlist with retailer affiliate).
Consumer AI·Full-stack·Production ML
03Internal production toolInternal · Loblaw
BadgeForge
A badge brief in, a reviewed, deployment-ready email badge out.
No public URL — internal Loblaw tool
An intake-to-deployment platform for the promotional badges in Shoppers Drug Mart and Loblaw CRM emails. Every request moves through a two-phase workflow, Brief then Badge Build, across submission, review, QA and sign-off, with role-based access for marketers, agency producers and admins over Microsoft Entra SSO.
It generates PDF briefs, handles bilingual EN and FR copy, builds UTM links, and validates subject lines with AI (Gemini 2.5 Flash). Transactional email fires at each stage and scheduled jobs send the reminders: daily digest, link-plan and UTM nudges, rejection follow-ups. Next.js App Router on Firebase App Hosting with Firestore and Storage. Solo build at LA Digital.
Internal tooling·Workflow automation·Next.js + Firebase
How I workProduction first
I start from the output a team actually has to ship — a planogram angle, a bilingual tile, a signed-off badge — and build backwards to the model.I start from the output a team has to ship and build backwards to the model.
Gates, not vibes
Every pipeline has a quality gate and a cost ceiling. Bad generations get caught before a user or a reviewer sees them.Every pipeline has a quality gate and a cost ceiling.
Solo to shipped
Design, build, deploy, support. All three projects here went from idea to live users without a hand-off, including one through enterprise SSO and review.Design, build, deploy, support — no hand-off.