Case Studies

Six case studies grounded in real products I've built. The seventh is this site itself. Each demonstrates a specific production AI engineering pattern with eval data, cost analysis, and honest failure-mode notes.

Case Study #1planned

Autonomous Trading with Safety-First Architecture

WealthPilot's 10-layer risk control architecture, regime-aware kill-switch, and weekly walk-forward retrain — the patterns a hedge fund customer needs to actually deploy LLM-driven trading agents.

safetyagentstradingevals

Target: 2026-08-15

Case Study #2planned

Multi-Agent System with Role Separation

Hermes Agent System — permission-layer separation of duties across Midas, Oracle, Plutus, and Kuber. How to give an enterprise multiple autonomous AI operators without them stepping on each other.

agentsmulti-agentobservability

Target: 2026-10-15

Case Study #3planned

Production Vertical AI in a Regulated Domain

LastingPath + Sage (Claude reasoning layer) — structured workflow routing, OCR-driven form auto-population, audit logging. What it takes to ship consumer AI in estate, tax, or healthcare.

vertical-aistructured-outputcompliance

Target: 2026-11-15

Case Study #4planned

Domain-Specific Document Understanding at Scale

TaxEXP — K1 and Form 1065 generation pipeline with extraction validation, confidence scoring, and human-in-the-loop fallback. Cost-modeled across Haiku/Sonnet/Opus.

document-understandingstructured-outputcost

Target: 2026-12-15

Case Study #5planned

Custom Claude Skill + Agentic Workflow

Reel Generator — packaging a domain-specific creative workflow as a Claude Skill, with quality gates and brand-voice evals in the agentic pipeline.

claude-skillsagentsevals

Target: 2027-01-31

Case Study #6planned

Eval-Driven Hardening of an LLM Product

50-prompt adversarial dataset, regression eval suite, and a failure-mode taxonomy with detection-rate metrics — applied to Sage in LastingPath.

evalssafetyred-team

Target: 2027-02-28

Case Study #7planned

The Tapasya Tracker (Field Notes Itself)

$0-budget product engineering, schema-first design, and 12 months of dogfooding observations from running my own deliberate-practice tracker.

metaproduction-engobservability

Target: 2027-04-30