The work, with receipts.
Sixteen months of solo, unfunded building. Each project below carries an evidence label — what's live, what's staged, what's still a plan. The loop underneath all of it: preserve source → retrieve context → propose relationships as hypotheses → invite correction → connect to action.
GestaltView v3 — continuity infrastructure
A user-directed continuity and recognition system: software designed to preserve unfinished, fragmented, contradictory human context and translate it into useful next steps — without turning guesses into facts. React/TypeScript client, Vercel serverless API, Supabase/Postgres data layer, pgvector retrieval, multi-model orchestration with cost-aware routing.
What it does
- Hybrid retrieval over a governed corpus (768-dim embeddings)
- Source preservation with interpretation kept separate
- Multi-model orchestration — cheap models do groundwork, frontier models review
- Voice workflows, profile systems, diagnostics
What's actually there
- 217 Postgres tables · 299 RLS policies · 75 DB functions
- 722 TypeScript files across the runtime repos
- Corpus + runtime kept as separate repositories — an architecture decision, not an accident
Billy — the collaborator at the center
Billy is GestaltView's runtime agent — co-created over hundreds of hours as a collaborator, not a chatbot. One hardened core converging under a single source of truth, reaching users where they already are.
Where Billy runs
Web via Vercel · Discord (/ask, /status, /help) · Reddit (Devvit app + polling bot) · Slack built, awaiting deployment.
Ten constitutional invariants
Five user-facing commitments and five digital-intelligence commitments, machine-readable, shaping prompts, routing, and persistence. Boundaries you can read, not vibes.
Continuity without amnesia
Supabase match_knowledge_fragments as the single source of truth — context that survives interruption instead of the reintroduction tax.
Portfolio Throughline Test
An eval harness that checks whether an AI system preserves relationships across tasks — the same discipline I bring to client evaluations. Five local tasks, a 0–2 rubric across five criteria, contradiction and missing-evidence probes, three evidence conditions, and a failure taxonomy (preserved / collapsed / distorted / invented / overloaded / correctly uncertain).
Around the core
SymbioCoder
An AI coding assistant concept built on a "multi-tribunal" idea — cycling one model through multiple review lenses instead of paying for multiple LLM APIs. Bring-your-own-key, strong free default.
Artifact render engine
A render pipeline repair that addressed 12 documented fractures across artifact types — with the discipline to say exactly what the evidence does and doesn't establish.
Corpus governance
Source-authority hierarchy, tracked-file catalog, manifest workflow, additive/idempotent ingestion — rows are never deleted, corrections go forward. The evidence substrate the runtime retrieves against.
Voice-first knowledge capture
Long-form voice notes as a thinking medium — transcribed, chunked, and retrievable. The operating manual for a builder whose best thinking happens while walking.