I build practical AI systems that automate real business workflows.

01 · working software

AI Support Approval Desk

Support tickets answered by AI, approved by humans.

Problem

AI can draft most support replies, but one wrong promise or a tone-deaf reply to an angry customer is a real business risk. The goal isn’t autonomy, it’s safe speed.

What I built

A reviewer console where tickets arrive via webhooks or email (n8n), get classified and drafted against a citable policy pack, then wait. The reviewer sees the draft, its policy sources, model telemetry and a full audit trail, and approves with a passcode before anything reaches a customer.

Outcome

• Every draft grounded in citable policy, with sources shown right next to the reply
• 10/10 eval cases passing across classification, sentiment, priority and safety
• Drafts arrive in ~3 seconds, with model, token and cost telemetry on every run

• Next.js
• TypeScript
• Prisma
• Supabase
• n8n
• Claude / GLM
• Resend

support-desk / reviewer-workbench

02 · working software

Document-To-Action Pipeline

Meeting notes in, evidence-backed tasks out.

Problem

Decisions live in transcripts, notes and email threads. Turning them into assigned, deadlined work is slow. Worse, owners get guessed and nobody can verify why a task exists.

What I built

A pipeline that reads documents and extracts decisions and action items under one hard rule: no quote, no export. Anything missing an owner or deadline is held for review, and only human-approved tasks are exported to Notion or Slack.

Outcome

• ~44 minutes / ~$44 of estimated ops value per document run, for $0.0003 of AI cost (verified live run)
• Every exported task traces to an exact source quote with a confidence score
• Suspected duplicates and unowned tasks are held for a human, not exported

• Next.js
• TypeScript
• Prisma
• PostgreSQL
• Clerk
• DeepSeek
• Notion & Slack APIs

document-pipeline / approval-queue

03 · working software

Lead Response & Booking Center

First to respond wins the job.

Problem

Service businesses lose work by replying late, but nobody wants an AI texting their customers unsupervised.

What I built

An intake center that catches every inbound lead, scores it 0–100 with readable reasoning, drafts the reply and can offer a real timezone-aware booking slot. Nothing reaches a customer until a person clicks approve. The gate is enforced in the architecture, not just the UI.

Outcome

• Every lead scored 0–100 and qualified hot / warm / cold / spam, with the reasoning on screen
• A ready-to-send reply in minutes instead of hours
• Six-stage CRM pipeline plus an audit log of every automated step

• Next.js
• TypeScript
• Prisma
• Supabase
• OpenAI
• shadcn/ui

lead-center / inbox

04 · in production use

AI Product Photography Studio

Campaign photos that don’t look AI-generated.

Problem

Product photos are expensive to reshoot for every angle, season and SKU, and naive AI generation produces warped labels and a different look every time.

What I built

A studio that enforces photography-grade prompt discipline before anything is generated. A prompt linter blocks lazy briefs, a Director workflow expands them into shot-ready specs, cheap draft takes come first, a human picks the winner, and follow-up scenes anchor to it for campaign consistency.

Outcome

• Three draft takes for $0.20: pick a winner before paying for finals
• Prompt linter in action: 12 errors blocked → 0 after the Director rewrite
• In production use as the engine behind Mira Content Studio

• TypeScript
• Node.js
• React
• Vite
• Gemini image APIs

mira studio / verre-noir · winter ritual

1. 01

   Map the workflow

   Before any code: what comes in, what needs to happen, where the risk lives, and what done looks like. If AI doesn’t earn its place in the flow, it doesn’t go in.

2. 02

   Build the system

   An app, an automation or an agent: small, typed, tested pieces on boring, reliable infrastructure. Working software over impressive demos.

3. 03

   Keep humans in the loop

   Approval gates on anything that touches customers, money or public output. Every automated step is logged and auditable.

4. 04

   Hand it over properly

   Documentation, runbooks, tests and cost telemetry. You get a system your team can run, understand and change without me.

Languages

• Python
• TypeScript
• JavaScript
• Go
• SQL

AI systems

• Claude API
• OpenAI
• Gemini
• MCP
• RAG
• Tool calling
• Model routing
• Evals

Backend & apps

• Next.js
• React
• FastAPI
• Node.js
• REST APIs
• Webhooks
• Prisma
• Supabase Postgres

Automation

• n8n
• Telegram Bot API
• Google APIs
• Notion API
• Resend
• Playwright

Infrastructure

• Docker
• AWS Lightsail
• Linux
• nginx
• Cloudflare
• systemd
• GitHub Actions

1. 2026 · now

   AI Automation Developer · Mira Solutions · London

   Freelance builds: AI agents, internal tools and workflow automations shipped to production, with approvals, observability and handover built in.

2. 2023 · 26

   Founder & Operations Manager · MIRA Properties · London

   Ran a short-let property business end to end, and automated the repetitive half of it with n8n and Claude workflows.

3. 2022

   Internal Operations Associate · Datto · London

   IT operations at scale: high-volume support across walk-up, phone and remote channels, plus a Google Workspace to Microsoft 365 migration.

4. 2021 · 22

   Business Operations · Motherland Organics · London

   Multi-channel e-commerce operations across Shopify, Amazon and social commerce.

5. 2019 · 21

   Assessment Operations Administrator · AAT · London

   Regulated, GDPR-sensitive process work: appeals handling where accuracy and auditability were the job.

BA (Hons) Business Studies · University of Greenwich