Case Studies & Projects 

Operate a complete commercial AI product with six subscription tiers and daily AI inference without runaway costs or manual operations.

Built on Supabase with Stripe billing, an n8n-orchestrated daily inference engine with a hard per-run cost ceiling, and a React Native app on Expo for iOS and Android with push notifications tied to daily content drops.

• Live with paying subscribers
• Six subscription tiers served from one engine
• Scheduled daily inference with capped unit economics
• One mobile codebase shipped to both app stores

NGOs lose weeks finding the right grants and drafting applications, and most can't afford a dedicated grants officer.

A platform that ingests grants daily from Grants.gov, EU Funding & Tenders, and UKRI, matches them to the NGO's profile, and drafts applications with Claude for editing and PDF export — with zero manual upkeep on the database.

• Live at grantifier.co.uk
• Daily automated grant ingestion with no manual upkeep
• AI matching and application drafting in production

Launching a fashion brand normally means recurring photoshoot costs for every collection — model fees, studio time, and weeks of lead time before a product can go live.

Built the storefront and a custom Flux LoRA brand model that generates editorial-grade campaign and product imagery on demand, keeping every drop visually consistent at zero studio cost.

• Live at shop.marymia.co.uk
• Custom AI brand model produces consistent editorial imagery
• New collections photographed and listed without studio costs

The provider needed a credible, modern site their team could update without breaking production — and a migration that wouldn't interrupt the email their operations run on.

A Next.js site on Vercel with automated deployments, split-routing DNS (web traffic to Vercel, email kept on the existing mail host), and a locked asset pipeline so photo swaps are filename-only changes.

• Live in production
• Zero-downtime DNS cutover — email untouched
• Client-safe content workflow, no code edits needed

Schools often can't see student disengagement until grades slip — the signals live in separate systems nobody checks daily.

A gamified engagement platform on Next.js and Supabase that unifies the signals into one staff dashboard with points and streak mechanics, deployed for a Washington, DC high school.

• Live and in daily use
• Engagement visible and actionable in a single dashboard
• Gamified mechanics driving participation

Run a daily content operation — scripting, voice synthesis, video rendering, packaging, distribution — without dedicating any headcount to it.

Claude drafts the script against an editorial standard, ElevenLabs synthesizes narration, Remotion renders the video programmatically, and n8n orchestrates packaging and scheduled publishing with mobile approval gates and automatic failure recovery.

• In production with scheduled daily output
• The work of a small production team at API unit cost
• Audit trail on every run — inputs, decisions, actions, cost

Outbound pipelines starve without a steady source of qualified leads — and bought lists are stale, non-compliant, and rented rather than owned.

Automated harvesting of advertiser leads from trade publications with enrichment, dedupe, and scoring against the ideal customer profile, feeding compliant outreach queues. Proven on our own pipeline before any client deployment.

• 380+ qualified leads from a single automated sweep
• CAN-SPAM-compliant outreach queues for US campaigns
• The same system runs our own outreach today

Manual cloud posture reviews take days, vary between reviewers, and produce evidence that rarely survives contact with an auditor.

Scripted checks built on Terraform inventory and Python/boto3 that write findings directly into our C/H/M/L register format — repeatable on demand after every remediation.

• A day of manual cloud review compressed into minutes
• Findings land in an auditor-usable severity-coded register
• Repeatable on demand

A HIPAA-regulated operator asked what AI could safely automate across intake, scheduling, compliance documentation, and family communication.

An eight-agent design with strict PHI boundaries, human approval gates on every agent, and full compliance logging — scoped as a fixed-price starter package with measurable hour-savings targets per workflow.

• Full written architecture and fixed-scope proposal
• 100% of agent actions logged and human-gated
• A worked example of scoping agentic AI in regulated environments