AI
Systems.

Agency work creates constant demand for strong first drafts across email, banners, ad creative, landing pages, and social copy — with zero tolerance for voice drift. Generic AI output is the enemy. Volume without precision just produces more wrong work faster.

Brand context files, RAG-retrieved voice samples, SKILL.md governance, and persona specifications are assembled before any draft is generated. The system produces multiple first drafts, then routes them through a standards AI for quality and voice checks before anything reaches a human for approval.

Context ingestion → multi-draft generation → standards-AI critique loop → Telegram approval routing → human edit and publish. Outputs span email campaigns, banner copy, landing pages, ad creative, campaign concepts, and social posts. Nothing ships without human sign-off.

Sunset Marquis: high-frequency promotional and email copy inside a brand voice that took years to build. Agency689: proposals generated from discovery notes and project requirements, structured against years of precedent. Traveler Guitar: systems thinking extended into visual ideation — creating campaign worlds where the product lives without losing the brand's sense of adventure.

Used daily across Agency689 client work. The practical version of AI-assisted creative: faster iteration, tighter review loops, and human approval before anything publishes.

Markets generate data at a pace no individual analyst can track. EDGAR filings, earnings transcripts, analyst revisions, macroeconomic feeds, and price signals all move simultaneously. Manual synthesis lags. Bad signals with high weights cause outsized damage. A system without governance is worse than no system at all.

Node.js + SQLite backbone. Claude Haiku handles routine daily reviews; Claude Sonnet handles advanced analysis. Ingests Finnhub, EDGAR, earnings transcripts, analyst revisions, and macro feeds daily. Claude synthesizes everything into structured intelligence briefs. Signal weights are automatically adjusted based on observed hit rates over time.

Every signal change moves through five verification states: UNVERIFIED → SYNTAX_OK → SQL_VERIFIED → RUNTIME_VERIFIED → PRODUCTION_VERIFIED. New signals receive 0.50× weight until proven. Sub-55% hit rates (n=15) trigger automatic deactivation. A 20% drawdown fires a system-wide kill-switch. Nothing trades automatically.

Launchd scheduling on macOS. Historical backfill from 2022-present (162,439+ price rows). Six hypothesis tiers: fundamentals, peer dynamics, AI infrastructure, VRT-specific factors, macroeconomics, and emerging research. COMPOSITE_BULL averages +3.65% return over 3-day holds. Everything is logged, versioned, and auditable. The changelog reads like a system's operating history.

The transferable design pattern: ingestion, hypothesis testing, composite scoring, human review, audit trail, continuous improvement. The SSIA white paper documents the full architecture.

Indie authors need consistent multi-channel presence. Platform-native copy for Reddit, X, Instagram, Facebook, Goodreads, and KDP all have different voice expectations, character limits, and community norms. Maintaining all six manually is a second job. Generic AI copy, posted without human review, does more damage than nothing.

A single Claude brain reads book metadata — title, author, target audience, ASIN — and generates platform-specific copy for all six channels in one run. OAuth2 and OAuth 1.0a handle platform authentication server-side. No credentials hit the frontend. KDP ranking data feeds back in via Rainforest API or CSV upload for attribution learning.

Local Node.js server with a browser dashboard. Single HTML frontend, single server backend — no build step required. Server health checks show which channels are live. Attribution tracking identifies which content and which channels convert to sales, then feeds those learnings back into the next generation cycle.

BookLite is the open-source Lite release of the internal Agentic689 book and music marketing system. It runs one book, one author voice, one Claude brain, and posts to six platforms. The full production version handles multiple titles, multi-artist configurations, and cross-platform attribution at scale. The Lite release strips to the essentials so the architecture is readable.

Built on the Anthropic Claude Agent SDK. MIT licensed. The white paper covers the full internal architecture the Lite version was derived from, including the multi-title production system.

Traditional mineral exploration is expensive, slow, and requires specialist teams in the field. Satellite imagery covering every inch of Earth's surface is freely accessible but opaque to non-specialists. The gap between available data and actionable prospectivity intelligence is a systems problem — not a data problem.

AuScan trains Claude Vision on the spectral and structural signatures of ten world-class gold deposits — Witwatersrand, Carlin Trend, Kalgoorlie, and others. Multi-sensor data — NASA EMIT hyperspectral, Landsat thermal, ASTER mineral mapping, Sentinel-2 vegetation indices, and ESRI RGB — is assembled into terrain fingerprints for each deposit.

In testing, the system scored a HIGH-confidence hit at 42.609°N, −119.299°W — Oregon's Pueblo Mining District, a documented former gold producer. Cross-referenced against the USGS Mineral Resources Data System. Search grids are configurable from 2×2 to 5×5 across 10–100 mile radii. Patterns and results persist across sessions.

React-based browser application running in a Claude.ai artifact environment. Single-endpoint design — no CORS, no separate auth layer. Browser-native storage for session persistence. USGS MRDS integration via Claude's knowledge base for cross-referenced validation. Configurable search parameters allow rapid iteration across different regions and deposit types.

The transferable insight: domain expertise can be distilled into a vision system that scales to any geographic target. The white paper covers the full methodology, sensor pipeline, and validation results.