Five projects, one argument: AI earns its keep when it ships and gets adopted by the people who
depend on it. Every project here maps to at least one of four pillars: delivery is the hard part;
build for the person depending on the system; logistics and sustainment are where the unglamorous
value lives; enterprise discipline makes AI safe to ship.
Accepted · Eval 24243a5a
Anchor · Pillars: Delivery · Operator · Logistics · Rigor
SustainAI: Predictive Sustainment & Supply-Exception Triage
Built for the problem I lived as a Marine logistician: too many exceptions, too little time
to find the ones that actually threaten the mission.
An RUL model predicts turbofan degradation on NASA C-MAPSS data, deterministic thresholds
band each exception by severity, and a single guardrailed Gemini agent triages every one:
0.842 critical-class recall at 0.941 precision (eval 24243a5a, reproduced bit-identically),
33/33 live triages on Vertex AI with 100% schema validity (eval 3477d2bf). The 3 missed
criticals are itemized by unit in the repo. Built by an autonomous coding agent through six
adversarial audit rounds; the eval harness, not the agent's self-reports, made the numbers
trustworthy.
Predictive Maintenance, Agentic Triage, Guardrails, NASA C-MAPSS, Evaluation Harness, Vertex AI
Shipped Product
Pillars: Delivery · Rigor
Teach-by-Doing V6
The production workflow-intelligence engine behind FreeFuse: it turns expert doing into
machine-readable workflows for humans, AI agents, and robots.
Matured across six iterative versions, from local MVP to a unified cloud engine, adding
hooks for compliance frameworks such as AS9100-style traceability along the way: the
discipline aerospace and defense manufacturing runs on. Live demo and public repo.
Workflow Intelligence, Robotics-Ready Knowledge, LSTM Sequencing, Pub/Sub, GCS, AS9100-Style Traceability
Built
Pillars: Delivery · Logistics · Rigor
ReneWind: Predictive Maintenance
Failure prediction tuned for the economics an operator actually faces: a missed failure
costs a replacement, a false alarm costs an inspection.
A tuned neural network classifier reaching ~88% recall on failures on unseen data across
25,000 labeled observations, with the decision threshold optimized to ~0.29 against a
custom cost function. The proven pattern SustainAI extends into sustainment.
Predictive Maintenance, Cost-Sensitive Classification, Threshold Optimization, Explainability
Built
Pillars: Delivery · Rigor
FraudShield V3: Real-Time Anomaly Detection
Exception detection for operations: an anomalous transaction and an anomalous shipment
are the same problem shape, and this is the architecture that catches them in real time.
Built end-to-end as a multi-tenant fraud detection platform: streaming features, hybrid
XGBoost and Isolation Forest models, SHAP explanations, drift monitoring, and Terraform
infrastructure. Deployment, evaluation, monitoring, and production-readiness in one system.
Anomaly Detection, Real-Time Streaming, Hybrid Models, SHAP, Drift Monitoring, Terraform
Built
Pillars: Delivery · Operator · Rigor
DemoRAG V5: Trusted Knowledge for Operators
A production retrieval-augmented generation service built for one job: put trusted, cited,
source-grounded answers in front of the person who needs them in the field.
Async fan-out pipeline on GCP with Firestore vector search and a dual-write storage-decoupling
design (Firestore plus Parquet) that keeps retrieval fast and the source of truth durable.
Built on a medical-education corpus; the pattern is SOPs, maintenance manuals, and policy.
RAG, Firestore Vector Search, Dual-Write Storage Decoupling, Cloud Run, Terraform, Citations
