Skills and evidence

The skills, and the projects that prove them.

Nothing here is self-rated. A skill is listed because it appears in a project's real stack, and it links back to that project. The AI patterns show how I used models, not only which ones.

AI engineering patterns

How I use models, not just which ones

Nine patterns that recur across the projects. Each example links to the case study where it is built.

Structured JSON output with a deterministic fallback

Models are asked for strict JSON. The result is parsed and checked, and when it fails, plain code takes over.

  • Personal AI Job HunterCover letters and answers come back as JSON with exact keys. A template fallback is used when the AI fails, and the result is recorded as ai_with_fallback.
  • AI HypnotherapistEvery analytical step returns structured output validated against Pydantic models.
  • GenbooksChapter plans are strict JSON. When a customer supplies an outline, the model must follow its exact titles and order.
  • Backlink BuilderEach site judgement is JSON with a verified quote behind every claim.

Model failover, retries and rate-limit handling

Free tiers and busy APIs fail. The client is built to carry on.

  • Backlink BuilderFour-model Gemini failover, several keys, a 5-minute budget per site and a stub client that answers 'busy' instantly after the first outage.
  • Personal AI Job HunterA shared limiter holds Gemini to 80% of the per-minute limit and honours Retry-After on HTTP 429.
  • GenbooksRetries on rate-limit and server errors using Google's suggested wait, aborts after 150 seconds of silence, and treats cut-off output as a failure.

Retrieval, chosen on purpose

Embeddings where they help, plain keyword search where they would only add risk.

  • AI Sales ManagerLangChain splitter, OpenAI embeddings and a saved vector index with a relevance cut-off. Approved documents go in directly when they fit under 6,000 tokens.
  • AI HypnotherapistIn-house BM25 with stemming and no embeddings. Unsuitable notes are excluded before indexing and unknown files are blocked.

Real-time and local voice

Two different routes to a talking product: a hosted speech-to-speech model, and open models on my own server.

  • AI Sales ManagerOpenAI Realtime speech-to-speech with interruption, a fresh session per guest and a ready connection handed over in 36 ms.
  • AI HypnotherapistKyutai Pocket TTS in the clinician's recorded voice and faster-whisper, both on the owner's ARM server, so there are no per-minute cloud voice fees.

Safety layers and guardrails

Several independent checks, with defaults that fail safe.

  • AI HypnotherapistFree screening before payment, a keyword net plus an independent AI classifier on every message, and a fail-safe when the classifier is down.
  • AI Sales ManagerLocked rules in code (AI disclosure, no medical claims, crisis handling) that no setting can override. A test proves it.
  • SocialFieldRule-based text moderation that sees through letter disguises, plus Google Cloud Vision SafeSearch on images.

Cost tracking and control

Pay-per-token and pay-per-voice products need a meter.

  • AI Sales ManagerA per-response token and cost log, with spend limits ($10 a day, $100 a month) checked at every start.
  • GenbooksInput, output and thinking tokens and image counts recorded for every job, including failed calls.
  • AI HypnotherapistPer-call usage and cost estimates, prompt caching, and a cheaper Haiku model for bulk reading.
  • Personal AI Job HunterUsage tracked per provider, and AI rating capped to a maximum number of jobs per search.

Grounding: never invent facts

The model may only use what it was given, and code checks where it can.

  • GenbooksPrompt rules against invented facts, studies and quotes. The author bio is rewritten only from the customer's own words.
  • Personal AI Job HunterCover letters use only the candidate's real background, and private notes must never be mentioned.
  • Backlink BuilderA 'yes' must come with an exact quote, and code confirms the quote is on the page.

Orchestrated pipelines, not free-roaming agents

Deterministic code decides the order. Models decide only what they are good at.

  • GenbooksOutline, front matter, preview, drafting, editing, repair and cover as separate calls, each step under a transactional lease.
  • Backlink BuilderA state machine allows only listed transitions and records the reason for each.
  • AI Sales ManagerA deterministic state machine. The model only decides whether to call one of three tools.
  • AI HypnotherapistSpecialised LLM calls with deterministic orchestration.

An agentic development workflow

Building with AI is itself an engineering discipline: roles, checklists and guardrails around the assistant.

  • SocialFieldFive custom Claude Code review subagents (payment, security, mobile, SEO, analytics) with diagnostic-only tools, carrying the project's own past bugs as checklists.
  • SinglesClaude Code sub-agents for research, security review and front-end work, plus GitHub Copilot.
  • GenbooksBegan as a human-supervised Claude Code book-writing workflow with sub-agents. The Gemini pipeline is a port of it.

Technical skills

Everything I have shipped with, and where

Built automatically from the stack listed in each case study. The number is how many projects use it. Pick a project to see only its skills.

AI models and APIs

AI techniques

Languages

Backend and APIs

Frontend

Audio and media

Data and storage

Payments and integrations

Cloud and infrastructure

Automation

Testing and tooling