Work

• Flagship

  Governed AI action platform: architecture, data platform and advisory agents

  Took an online retailer from a map of recurring departmental work to one shared, governed AI platform: rate-limit-safe nightly ingest into one warehouse, advisory agents that prove every figure and wait for a person, and costed designs for the next agents.

  Part design, not built

• Flagship

  Department AI agents with human approval gates

  A family of chat-based department agents on shared, hardened foundations where the model does the routine work but every consequential action waits for an expiring, single-use human approval and every figure comes from code.

• Case study

  Governed action pattern for an AI customer-support agent

  A fail-closed pattern that lets an AI customer-support agent take a high-risk, hard-to-undo action: the model only extracts, code authorises, one gateway acts through a write-ahead ledger, and a replay of past decisions before launch reports wrong actions apart from escalations.

• Case study

  secure-minimal-agent

  A small, public, provider-agnostic reference agent that shows security and correctness disciplines in as little code as possible, with its safety claims pinned down by tests.

• Case study

  Safe AI access to business data (MCP servers)

  Model Context Protocol servers that give AI assistants compact, cost-capped, read-only-by-construction access to business data and refuse rather than guess.

More work

• AI work-prioritisation and status planner

  Ranks all open work transparently, turns the ranking into protected calendar time, and generates status updates and meeting decks on demand and on a schedule.

  →

• Case study

  AI receptionists for SMS, voice and chat with code-checked booking

  AI receptionists that take an enquiry from first message to a confirmed appointment over SMS, voice or web chat, with code re-checking every slot the model proposes and no lead ever dropped, plus a labelled design for a safety-first phone receptionist in a regulated setting.

  Part design, not built

• Case study

  Realtime live-chat agent console with an AI reply assistant

  A realtime workspace for support agents handling many chats at once, with visitor context, escalating wait reminders, human-approved and rated AI reply suggestions, and automatic recovery of missed messages when a laptop wakes from sleep.

• Case study

  Support desk automation: knowledge sync, in-chat booking and customer context

  Keeps an AI customer-support agent's knowledge current without ever destroying content, lets customers book inside the chat, and shows each customer's account context in the help desk the moment a conversation opens.

• Case study

  AI meeting notes, recaps and operations companion

  Captures meetings automatically, delivers structured recaps to the right place even when transcripts arrive late, lets each person approve their own action items, and can keep audio on company hardware.

• Case study

  Brand-safe generative media: AI makes the scene, code makes it exact

  Generative image, video and copy tooling where AI makes the scene but code applies everything that must be exact (typography, approved claims, mandatory wording), with people signing off at each decision.

• Case study

  Self-hosted AI tutor for a family, with code-enforced safety rules

  A self-hosted AI tutor for children where the safety rules are enforced in code rather than the prompt, facts are kept correct where they matter, and quality is measured over time.

• Case study

  Automated water test bench prototype with image classification

  An inexpensive bench-top prototype that automates water sampling, chemistry readings and sample preparation for imaging, sorts the prepared sample with an image classifier and merges both results into one overall risk rating in plain words.

More work

• Local multilingual video and live-stream summarizer

  Turns a video or live-stream recording into a short English summary produced mostly on the local machine, flagging uncertain names and stating unclear stretches instead of papering over them.

  →

• Recovery-aware AI workout planner

  Builds each workout from per-muscle recovery, available equipment and a time budget, with instant offline logging, the AI choosing and explaining while code owns the arithmetic.

  →

• Case study

  Evidence-first workflow skills for an AI coding agent

  Repeatable commands that make an AI coding assistant work like a careful lead engineer: find where things really stand, bring in independent second opinions, keep every quote and number traceable, and stop for a human at each decision.

• Case study

  Governed AI development workstations

  One idempotent install that gives every workstation the same governed AI setup, with credential blocking and redacted audit that fail closed and a verifier that attacks its own controls.

• Case study

  Gamified AI-adoption leaderboard

  Automatically measures AI-assistant usage with no effort from staff and turns it into friendly recognition through a leaderboard and chat highlights, with company-only sign-in and role-protected data.

More work

• Allowlist-gated browser harness for testing conversational AI agents

  A harness that lets a coding agent test a conversational AI agent one message at a time, with the safety policy enforced in code so it cannot touch anything not explicitly allowed.

  →

• Zero-friction time tracking and guard rails for AI coding sessions

  Records engaged time per task automatically from AI coding sessions, never counting idle time or storing prompt content, and puts declarative guard rails around what the assistant may do.

  →

• Case study

  One customer across many systems: billing and identity integration

  Keeps one trustworthy customer identity in step across billing, website and support systems, with confidence-ranked matching, human corrections that stick, failures that reach people, and self-service billing where no card data touches the site.

• Case study

  Multi-client marketing analytics warehouse and report generator

  One scheduled platform that pulls every marketing source for every client, models consistent monthly metrics behind quality gates, generates each client's monthly deck and proves the dashboard still matches its specification.

  Part design, not built

• Case study

  Resumable multi-source sync engine for project analytics

  Pluggable connectors pull project and time data from many SaaS tools into one store with checkpointed, resumable runs, locking with stale-lock recovery, deduplication and identity resolution.

More work

• Subscription churn and renewal analytics pipeline

  Rebuilds renewal history that the billing system overwrites and turns it into trustworthy monthly churn and renewal metrics in a secure dashboard.

  →

• Web traffic forensics and website audit pipeline

  A reproducible pipeline that turns raw web logs and read-only site access into a defensible traffic breakdown and a prioritised health and security audit, with evidence behind every conclusion.

  →

• IT and identity operations automation

  Turns recurring IT chores into safe tools: one-command credential rotation across many services with a secret-free record, a signed, access-controlled real-time message inbox, and read-only mail-delivery investigations.

  →

• Cloud environments on demand: disposable training labs and isolated per-customer hosting

  Repeatable cloud environments that are isolated, reachable without public IP addresses, cheap between uses and safe through every lifecycle step.

  Part design, not built

  →

• Safe follow-up assistant for unanswered emails

  Finds sent emails still unanswered after a set number of days and sends polite follow-ups inside the original thread, with the user controlling every message and no thread offered twice.

  →

• Rate-limit-aware unattended batch automation CLI

  A packaged command-line tool that runs a daily batch unattended, never acts twice on the same item, respects a service's rate limits with its own conservative request budget and backs off when throttled, and verifies afterwards that its work registered.

  →

• Self-auditing household budget workbook

  A spreadsheet that builds, checks and repairs itself, logs scheduled debits on its own and says how much to move where each pay cycle, with no external service.

  →

• Flagship

  Crew and equipment scheduling: from discovery to production

  From interview- and workbook-based discovery and a priced build-versus-buy decision to a production scheduling platform whose tested conflict engine catches clashes as the plan is edited, with enterprise sign-in and managed database operations.

• Case study

  Self-service portals: member networking and employee HR

  Two self-service portals, one for a professional networking group's members and one for a company's employees, sharing role-aware views, admin view-as-user, notifications and live integration with back-office systems.

• Case study

  Multiplayer garden game: from prototype to realtime serverless rebuild

  A multiplayer browser game built twice: first with a server-authoritative economy on plain shared hosting, then rebuilt on a serverless stack with realtime rooms, managed sign-in and scheduled world ticks.

• Case study

  Hyperlocal rain alert menu bar app

  A free, keyless desktop menu bar app that warns about rain at four speeds, states where every number comes from and how certain it is, and learns its own accuracy over time.

• Case study

  Self-running daily reading and reflection app (two versions)

  A daily reading and reflection app that assembles each day before anyone wakes, never shows half-built content, degrades visibly on failure, is designed around free-tier allowances, and was rebuilt when the first platform could no longer carry it.

More work

• WordPress and web engineering

  Web engineering inside the tools a business already uses: AI analytics explanations with capped spend, reversible bulk price edits, never-miss scheduled greetings, and a static site rebuild designed to break no ranked URL.

  →

• Map-based land listings site

  A map-based listings site where buyers browse real lot boundaries on satellite imagery, compare and inquire, and a listing agent publishes lots with drawn boundaries, document checklists and hazard notes.

  →

Earlier work

• Multi-vendor online shop with several payment methods (earlier work)

  An earlier multi-vendor online shop with seller shops, checkout with several payment methods confirmed by asynchronous callbacks, a wallet, two-factor sign-in, identity verification and seller reputation scores.

  →

• Case study

  Secure multi-role staff and client portal (architecture design)

  Architecture for a single sign-in staff and client portal that protects staff records and client data by structure: server-side authorisation, layered client isolation, a separate HR permission and an append-only audit log.

  Design, not built

More work

• AI-ready company knowledge: procedure library preparation, access design and retrieval rules

  Preparing a company procedure library to be safe and machine-readable for AI assistants, and a secure, auditable access and retrieval design where everyone reads, few write and access narrows by group membership.

  Part design, not built

  →

• Flagship

  Governed publishing pipeline and sanitized AI portfolio

  This portfolio: private work reaches the public only through a machine-enforced chain of sanitized draft, deterministic privacy scan, explicit human approval and a signed public record anyone can verify, with an assistant that answers only from approved content.