Case studies
Each case keeps its limitations visible: what was verified, what stayed uncertain, and where the boundaries held.
Portfolio map
Further cases are in preparation and appear here once their claims pass review.
01 · Agent orchestrationOperational
A repeatable, AI-assisted readiness workflow for SAP HANA analytics changes: automated multi-phase regression checks, specialist-reviewed findings, and controlled reporting — with production activation cadence rising during its staged adoption.
02 · Secure MCPOperational
A governed MCP boundary gave an AI assistant safe access to live network telemetry. Outlier analysis then turned monitoring data into bounded, verified fixes: a miswired uplink termination, a camera error path, and a least-privilege firewall change.
03 · Deterministic automationOperational
A deterministic climate controller with the governance visible: profile-based control with interlocks and fail-safes, a power-verified air-quality loop, and a joined sleep/environment analysis that declines to overclaim.
04 · Agent orchestrationOperational
The pipeline behind this site: an agentic team with bounded roles — builder, independent reviewer, interviewer, records clerk — deterministic privacy gates on every build, and human consent as the one deliberately unautomated step.
05 · Agent orchestrationOperational
Self-hosted personal-finance automation under a written operations contract: cost-tiered updater agents over MCP, human-gated imports and reconciliation, a ledger designed so tax season and life-to-date questions become filter operations, and every statement sanitized on the operator's own machine before any model reads it.
06 · Data governanceOperational
A self-hosted health data platform with no database port by design: three data streams landed provenance-first, every write dry-run and backup-gated, metric definitions pinned and test-enforced, and the most sensitive artifacts placed where the report web server cannot reach them.
07 · Agent orchestrationOperational
An agent team that plans my training, reads back what actually happened, and proposes changes — where judgment sits with one orchestrator, execution sits with cheap narrow specialists, and nothing reaches the device before I accept it.
