Applied AI Systems — Project Brief

[email protected] · (507) 208-0737 · linkedin.com/in/rshangle

This document describes two systems I designed and built independently. Both are production-operational. I am including this brief because technical success roles centered on AI adoption require more than a resume line — the design decisions here are the actual evidence.

Flatline Codex — Local-First AI Operations Kernel

A personal AI operations infrastructure built around a 105-table SQLite database (WAL mode) that functions as the canonical state for all system activity. The architecture prioritizes traceability and human-in-the-loop control above automation speed.

Key components:

Tiered lookup router (PYTHIA) — resolves queries through a deterministic chain: in-memory cache → indexed DB lookup → full-text search → semantic vector search (ChromaDB, 1,571 document chunks). Each tier logs source and latency. Misses are classified and tracked. The goal is measurable answer quality, not just answer speed.
Ingest pipeline (JAUNT) — content enters through a structured intake funnel with a mandatory human review gate before anything reaches the active corpus. At each stage, a disposition is logged. Nothing enters without a human accept decision.
Drift monitoring (SENTINEL) — seven nightly checks covering schema drift, index staleness, orphaned files, and stale worktrees. Results write to a drift_report table. Auto-remediation is scoped and logged; anything outside scope alerts for manual review.
Audit architecture — every significant agent action writes a breadcrumb to a message bus and an ops run log. The system can reconstruct what happened, when, and why for any point in its history.

JAUNT Intake Pipeline

Stage	Description
Auto-Triage	Incoming content classified by type, domain, and quality signal
Coordinate	Canonical coordinates assigned (domain, structure, purpose, terrain)
Validate	Frontmatter checked; schema enforced; malformed files routed to quarantine
Human Review	Accept / Pass / Trash — nothing proceeds without explicit human decision
Custody	Accepted files stamped, moved to active corpus, logged to custody chain

Why it's relevant: This is the operational pattern AI adoption in regulated environments requires — structured intake, human review gates, traceability at every layer, and a clear separation between "model suggested" and "human approved."

flowchart TB
	A["Flatline Codex<br>Local-First AI Operations Kernel"]:::title
	A --> B["Canonical state<br>SQLite (105 tables, WAL)"]:::core
	B --> C["PYTHIA<br>Tiered lookup router"]:::comp
	C --> D["JAUNT<br>Ingest pipeline"]:::comp
	D --> E["SENTINEL<br>Drift monitoring"]:::comp
	E --> F["Audit architecture<br>Message bus + ops run log"]:::comp

	C --> C1["In-memory cache"]:::tier
	C1 --> C2["Indexed DB lookup"]:::tier
	C2 --> C3["Full-text search"]:::tier
	C3 --> C4["Vector search<br>ChromaDB (1,571 chunks)"]:::tier
	C4 --> C5["Every tier logs source + latency<br>Misses classified + tracked"]:::note

	D --> D1["Auto-Triage"]:::stage
	D1 --> D2["Coordinate"]:::stage
	D2 --> D3["Validate"]:::stage
	D3 --> D4["Human Review<br>Accept / Pass / Trash"]:::gate
	D4 --> D5["Custody<br>Stamp + move + log"]:::stage

	E --> E1["Nightly checks (7)<br>schema drift • index staleness • orphans • stale worktrees"]:::stage
	E1 --> E2["Writes drift_report table<br>Scoped auto-remediation + alerts"]:::note

	classDef title fill:#0b1f3a,stroke:#0b1f3a,color:#ffffff;
	classDef core fill:#e8f0ff,stroke:#4c6fff,color:#0b1f3a;
	classDef comp fill:#f1f5f9,stroke:#94a3b8,color:#0f172a;
	classDef tier fill:#ffffff,stroke:#cbd5e1,color:#0f172a;
	classDef stage fill:#ffffff,stroke:#cbd5e1,color:#0f172a;
	classDef gate fill:#fff7ed,stroke:#fb923c,color:#7c2d12;
	classDef note fill:#f8fafc,stroke:#e2e8f0,color:#334155;

Dreadnaught — Modular Prompt Engineering System

A structured prompt assembly and evaluation system built to enforce reproducibility and source discipline.

Key design decisions: