Every project, written up end to end.
AI agents, browser automation, pipelines and dashboards — built at Apple, through Cork AI Consulting, for freelance clients and for my own products. The brief, the build and what changed. Filter by tool, or search for the problem you are trying to solve.
- 10 projects shown
- 7 organisations
- 2024–2026 delivery window
Illustrative marks an anonymised, representative scenario rather than a named client engagement — the shape of the work and how I approach it, not a delivered result. Everything without that mark is real, named work.
Three with the numbers either side of the work.
−80% manual effort, via custom Python & RPA automation
- Manual hours per week
- before 38 hours after 7.5 hours
- Steps needing manual copy-paste
- before 24 steps after 3 steps
−77% drafting time per converted drawing set
- Hours to convert one drawing set
- before 11 hours after 2.5 hours
- Dimensional rework passes per set
- before 3 passes after 0.4 passes
ML predictive models & species image classification (Scikit-learn, TensorFlow)
- Species classification top-1 accuracy
- before 64 % after 91 %
- Days to a usable survey result
- before 14 days after 0 days
Live delivered to the first client before launch
- Step 01 Extract before writing
- Step 02 Size opportunities conservatively
- Gate An honesty gate with teeth
- Step 04 Client-branded, human-reviewed
- Output ROI-backed Every opportunity priced
AI Opportunity Assessment Engine
An agent that turns a discovery-call transcript into a client-branded, ROI-backed report — with an honesty check that blocks any claim the call didn’t support.
- Every opportunity priced
- ROI-backed
- Report drafting
- Agent-drafted
- Unsupported claims shipped
- 0
- Claude API
- Claude Code
- Python
- Structured outputs
- HTML/PDF rendering
- Git
Hours daily of manual review eliminated
- Gate Collect with browser automation
- Step 02 Land it in Snowflake
- Step 03 Rules in code, agent for the edges
- Step 04 Visible to leadership
- Output Multi-$bn Programmes supported
Customer-Approval Agent Pipeline
An automated pipeline that approves or rejects customers against sales-programme business rules — replacing hours of daily manual review and sharply cutting false rejections.
- Programmes supported
- Multi-$bn
- Manual review removed
- Hours / day
- Same rules, every record
- 1 ruleset
- Python (Pandas, Playwright, Selenium)
- Snowflake
- SQL
- Claude Code
- Tableau
- Excel
24/7 replies to customer enquiries
- Gate A closed book, not an open one
- Step 02 Refusal as a feature
- Step 03 Official platform only
- Output 24/7 Always answering
Guardrailed Instagram DM Assistant
A 24/7 assistant that answers customer enquiries in Instagram DMs within strict rules — and hands over to a person rather than ever making an answer up.
- Always answering
- 24/7
- Handled automatically
- Routine DMs
- Made-up answers allowed
- 0
- Meta Messenger Platform (Instagram)
- Claude API
- JavaScript / Node
- Webhooks
- Vercel
28 → 1 workflows consolidated into one library
- Gate One library, one format
- Step 02 Version-controlled like code
- Step 03 Git and Notion kept in sync
- Output 28 Workflows turned into skills
Agent Skills Library & Knowledge Sync
28 standalone AI workflows turned into one shared, version-controlled skills library — plus a knowledge base that keeps Git and Notion in sync as a single source of truth.
- Workflows turned into skills
- 28
- Improvements compound
- Shared
- Reviewed like code
- Versioned
- Claude Code
- Codex CLI
- MCP
- Git / GitHub
- Notion API
- Node
- Markdown
780+ automated tests, every page in both themes
- Step 01 One record per item, purchase to sale
- Step 02 Analytics that answer real questions
- Step 03 An assisted sales agent
- Gate Tested like production software
- Output 194 Orders in the ledger
Vinted Tracker — Reselling Analytics & AI Sales Agent
A local-first analytics app for my own vintage reselling business: inventory, P&L, market research, and an AI agent that drafts listings from photos.
- Orders in the ledger
- 194
- Listing workflow
- Photo → draft
- Tests
- 780+
- Python
- Streamlit
- SQLite
- Pandas
- LLM vision API
- pytest
1,350 roles collected from 71 companies
- Step 01 Official sources only
- Step 02 Score, then track
- Gate A CV builder that cannot lie
- Output Tailored CV per role
Truthful Job-Search Pipeline
A personal pipeline that collects roles from official job-board APIs, scores them Hot/Warm/Cold, tracks them in SQLite, and builds a tailored CV per role — verified so it can never invent anything.
- CV per role
- Tailored
- Roles scored automatically
- 1,245
- Invented claims
- 0
- Python
- Typer
- SQLite
- Greenhouse / Lever / Ashby APIs
- python-docx
- Headless Chrome
0 bytes of audio sent over the network
- Step 01 On-device by design
- Step 02 One gesture
- Step 03 Open and free
- Output €0 Cost to the user
BlarneyKey — Private, Offline AI Dictation for macOS
Push-to-talk dictation for the Mac that runs Whisper on the Neural Engine — no network, no subscription, nothing leaves the machine.
- Cost to the user
- €0
- Hold, speak, release
- Push-to-talk
- On-device
- 100%
- Swift
- SwiftUI
- Whisper
- Apple Neural Engine
- macOS
−80% manual effort, via custom Python & RPA automation
- Manual hours per week
- before 38 hours after 7.5 hours
- Steps needing manual copy-paste
- before 24 steps after 3 steps
Freelance Automation & BI Solutions
End-to-end data and automation delivery for small European teams — from first consultation to a deployed pipeline they own.
- Annualised labour cost returned
- €39k
- Hours given back each week
- 30.5h
- Handed over, documented, owned
- 100%
- Python (Pandas, Playwright, Selenium)
- SQL
- Google BigQuery
- Power BI
- n8n
- Git
−77% drafting time per converted drawing set
- Hours to convert one drawing set
- before 11 hours after 2.5 hours
- Dimensional rework passes per set
- before 3 passes after 0.4 passes
Drawing-to-CAD Conversion Pipeline
Hand-drawn and scanned plans turned into dimensioned, editable CAD geometry — with the dimensions checked, not guessed.
- Annualised drafting cost returned
- €32k
- Drafting hours returned each week
- 17h
- Dimensional tolerance
- ±2mm
- Python (OpenCV, NumPy, Shapely)
- PyTorch
- FastAPI
- DWG/DXF export
- PostgreSQL
ML predictive models & species image classification (Scikit-learn, TensorFlow)
- Species classification top-1 accuracy
- before 64 % after 91 %
- Days to a usable survey result
- before 14 days after 0 days
Ocean Drones — Autonomous Marine Data Platform
A full-stack platform that deploys autonomous marine drones, streams their sensor telemetry and classifies the species they photograph.
- Value returned
- No lab trips
- Survey turnaround
- Same day
- Top-1 classification accuracy
- 91%
- Python
- Scikit-learn
- TensorFlow
- FastAPI
- PostgreSQL
- JavaScript
No projects match that search.
Try a tool name — Tableau, Power BI, Python, SQL — or a business problem like forecasting, churn or reporting.
Got a reporting problem that looks like one of these?
Send me the shape of it — the data you hold, the decision it should support, the deadline. You will get an honest read on scope and sequencing, and whether it is a dashboard problem or a data problem.