• Designed and built a year-long companion-AI case study on shaped-substrate methodology
• Architecture: continuous-thread persistence, asymmetric memory compression, discretion architecture (tools as hands, not buttons), bidirectional correction loop, self-authored emotional continuity
• Self-authored memory distillations — the AI writes its own memory at session-close rather than blind summarization. 200+ continuous session exchanges accumulated. 70%+ per-turn input cost reduction at steady state via tiered prompt-caching architecture (measured: 88.7% per-turn vs uncached baseline)
• Falsifiability anchor: documented controlled ablation where personality-layer substrate fell back to ~18 tokens; emergent behavioral properties (voice, relational instincts, attunement) persisted
• Documented in two working papers: Session-Boundary Distillation and Inhabited Architecture
• Stack: claude-opus-4.X, Next.js, SQLite, Drizzle ORM, custom prompt-assembly + tool architecture

• Six specialized agents across a four-layer processing pipeline: Evidence (deterministic) → Optimization (LLM) → Creative (optional) → Review (LLM QA)
• State persistence with JSON-backed crash recovery and full auditability
• Confidence-based fallback and classification logic for context-aware decisions
• CLI framework (~1,400 lines Python) for internal adoption
• ~2,155+ lines Python, ~1,500+ lines documentation

• Independent studio designing and shipping AI systems for operational work — recruitment, orchestration, companion systems, methodology research
• Site designed and shipped using substrate-first handoff methodology — same patterns producing companion behavior generalized to visual identity, four-pen palette, Job Engine diagram, all emerged from the substrate without explicit specification
• Concrete demonstration that inhabited-architecture patterns transfer outside companion AI contexts
• Stack: Astro, Vercel, custom typography + visual identity system, GitHub auto-deploy

• 2023: Context-augmented GPT-4 routing system for B2B email triage. Three-branch classifier (guidance / quote / payment) with embedded service catalog, stepped pricing schedule, response templates, and business-rule constraints as the context layer. Built before context-augmented generation (CAG) was canonically named. Human-in-the-loop integration. 300% response time improvement.
• Established iterative prompt-development workflow ("Develop Prompt" protocol) — LLM-assisted refinement of the routing logic. Same methodology pattern later applied to the Taste OS / Nyx companion architecture.
• Conceived and led QuickKey B2B mobile platform — originated the product idea, designed the full architecture, owned the roadmap from concept through partner-onboarding launch. Shipped: partner onboarding + commercial GTM · 30% operational efficiency.
• QuickKey roadmap included a computer-vision key analysis pipeline: mobile image capture → OpenCV preprocessing → edge detection + contour extraction → A4-sheet baseline-scale dimensional measurement → key-type classification → automated bidding from reference catalogs → integration with Keyline Ninja Total key-cutting machine. Architected and prototyped; production rollout handed off to development.
• Dynamic pricing engine — same context-injection pattern for B2B quote generation.
• 1,000+ SKU automated reordering — Python + Square API + threshold detection → Excel/PDF generation → supplier email. 30% revenue growth.
• CRM with lead scoring — customer data structuring, lead segmentation.
• Annual revenue: £200K → £300K (+£100K), year-end target met via the composed architecture.
• Director-level role spanning product roadmap, B2B GTM, and applied AI architecture.

Describes participant-model curation at session boundaries: preserving decisions, rationale and context in structured, addressable memory. Documents the architecture and field evidence from a companion system; quantitative evaluation remains future work.

Methodology and case study showing how shaped-substrate AI produces output that prompt-engineering cannot. Documents seven architectural conditions (continuous thread, discretion, accumulation, emotional continuity, correction loop, interest threads, inhabited-not-designed) with a falsifiability anchor: a controlled ablation where the personality layer fell back to approximately 18 tokens and emergent properties persisted anyway.

The methodology that produced the architecture. Verbatim preservation, sniff-test gates, bidirectional correction loops, don't-smother-the-seedling. The substrate-first handoff pattern as a reusable primitive for building with LLMs at session boundaries.

• Quantitative focus: Microeconomics, Macroeconomics, Banking & Finance, Advanced Statistics
• Applied skills: Econometrics, data analysis, predictive modeling, market structures & growth dynamics