He designs the system.The AI builds it.
Six case studies. Real systems. No computer science degree.
Every system here was built by defining the problem precisely, directing AI through a documented methodology, and evaluating the output against a measurable standard. The AI wrote the code. Thomas, an AI director, made every product and architectural decision. The six case studies span: an SEO content production system, a job fit evaluation system built on occupational psychology research, a full-stack lesson planning app for Ontario teachers, a file-based knowledge architecture running across eight active projects, a custom AI workflow tool built for managing concurrent AI sessions, and a voice capture system that learns writing style from edit behaviour.
See the WorkThe Work
File-Based RAG Architecture
Claude has no memory between sessions. Across 8+ active projects spanning months of work, every session was either context-heavy and slow, or context-light and error-prone.
A persistent, queryable knowledge base across 8+ AI projects using nothing but structured markdown. Four components: a structured index file per project (entry number, date, timestamp, summary, tags, chunk pointer), chunked log archives capped at ~1,300 lines, per-project tag vocabularies registered in CLAUDE.md, and per-project skill files that specify the exact logging format, keeping every entry consistent across sessions. No database. No vector store. No code.
The largest single project: 16 chunk files, 11,800+ lines of archived sessions. Zero database. Zero code.
Director · Architect · Zero code written
Read the case study →AI-Powered SEO Content Production System
Manual SERP analysis for a single keyword took hours to days. Scaling it was unthinkable.
A three-stage workflow: AI-directed domain research, encoding the research into a reusable prompt, then deployment against real SERP data. The system autonomously analyzed 8 competing pages, produced a full optimization blueprint, and generated a ~2,500-word article with FAQ content, schema markup, title and meta options, and link suggestions.
Same prompt run through Claude and Grok in parallel. Output quality was scored, and each gap drove a specific prompt revision.
Director · Architect · Zero code written
Read the case study →AI-Powered Job Fit Evaluation System
This site is a live output of CS2. The job postings that led here were screened through it.
Evaluating job postings manually was slow, inconsistent, and undocumented. No framework for distinguishing hard mismatches from closable gaps.
A three-phase evaluation system grounded in occupational psychology research. Phase 1 produces a role snapshot, fit score, alignment mapping, and divergence analysis (hard mismatches vs. closable gaps). Phase 2 runs an 8-signal environmental red flag screen. Phase 3 produces a structured summary for comparing postings side by side. Prompt refined through multi-model comparison (Fable, Opus, Sonnet) to a 95% benchmark match running at Sonnet cost. 40+ real postings evaluated.
Five benchmark refinement rounds: 80% → 85% → ~90% → 93% → 95% Fable-match quality, running at Sonnet cost.
Director · Architect · Zero code written
Read the case study →Ontario Teacher Unit Planner
Ontario teachers had two options: hours of manual curriculum cross-referencing, or fast AI tools that hallucinate curriculum expectations. A teacher who catches the error mid-unit has already built on a broken foundation.
A full-stack web app (Next.js, Supabase, n8n, Claude) with 14 active AI workflows, 13 Supabase tables, and all 3,717 Ontario Ministry curriculum expectations across all 10 subjects loaded as verified structured data. No curriculum content is ever generated by AI. Two-stage planning workflow: diagnostic-informed lesson skeletons reviewed before full expansion; sections like Teacher Notes and Differentiation are generated on demand. Dependency-aware revision cascade, curriculum coverage tracker, export to PDF/DOCX/Google Docs, asset generation infrastructure, and admin analytics dashboard with friction tracking.
Stage 2B context trimming: 14,750 → 4,832 input tokens per call (67% reduction). Four workflows migrated from premium model to Sonnet after cost audit found no quality justification.
Director · Architect · Zero lines of code written
Read the case study →AI Workflow Supervisor
Running 6–8 concurrent Claude Code sessions in a standard terminal is management-by-memory. No visual status when a session finishes, no project organisation, no persistence between restarts, and no way to see at a glance what is running and what needs attention.
A native macOS desktop app built entirely through directed AI development: session status detection using a timing-string pattern match in PTY output (an orange dot appears when Claude finishes — a feature that exists in no other terminal tool), session persistence across restarts, a project navigator with live session grouping, a file browser, and full keyboard navigation. ~2,300 lines of code across 11 modules. Zero lines written by Thomas.
The status indicator detects completion by matching Claude's timing string in PTY output — a pattern no existing terminal surfaces. ~2,300 lines of custom code across 11 modules, directed and delivered without writing a line of code.
Director · Architect · Zero code written
Read the case study →AI Voice Capture System
AI-assisted writing has a drift problem. Every session starts over. Style corrections from the previous session disappear, and the same generic habits resurface in each new draft. A style guide written from memory doesn't solve it — most voice characteristics are implicit, recognised in edits rather than described in advance.
A closed-loop voice capture system with four components: a pre-draft context filter (reads the active exchange to identify topic, addressee profile, and conversational situation, then filters the style ledger for matched signals); a three-agent adversarial review (Primary Drafter, Logic Auditor, and an Adversarial Reviewer whose persona shifts by topic type); style signal extraction from edits (reads user changes and extracts the underlying rule, tagged across three dimensions); and a tag-based style ledger with tiered promotion thresholds, phase-out logic, and a separate analogy repertoire.
62+ logged sessions. 75+ signals in the ledger. 20+ promoted to adopted status. 3 phased out, including one flagged as an AI tell and removed from use.
Director · Architect · Zero code written
Read the case study →One methodology. Six domains.
Every system here was built the same way: a research phase first, then encoding those constraints into a prompt or architecture, then deployment against real data and iterative refinement. Six case studies across six domains. The research phase changes. The framework doesn't. That's what makes this transferable, and that's what makes it worth hiring.