Portfolio · Selected evidence

Systems and decisions that can be examined.

Four projects built to address concrete problems. Each case presents the decision that shaped its design, the evidence available and the result’s actual scope. The complete activity remains on GitHub.

Customer Intelligence · Data governanceReference implementation

Data Intelligence Lab

A reproducible lab connecting a synthetic causal model with five Customer Intelligence cases and agents that consult metadata before analyzing data.

Context: Analytics and governance cases started from different assumptions, making it difficult to compare decisions and outcomes on a common foundation.
Design decision: A track-based monorepo: one fintech model feeds segmentation, churn, next-best-offer, ARPU and incrementality; governance agents read the catalog before touching data.
Evidence: Code, tests, synthetic data, project-level documentation and an integrated analysis of all five cases.
Outcome: Results change when the window, population, leakage boundary and causal answer are fixed before modeling: those choices are part of the system, not post-hoc cleanup.
Scope: Its purpose is to demonstrate reproducible decisions, not to operate as a product. The model retains documented simplifications around tariff migration and post-churn billing.

Code ↗Analysis →

Agentic AI · Governance and riskReference implementation

SigmaGuard

An experimental DeFi agent combining deterministic limits, statistical detection, AI explanation and human approval for anomalies.

Context: Automating financial decisions without separating limits, detection and approval makes it difficult to explain why an action executed or stopped.
Design decision: The architecture separates four responsibilities: deterministic rules, Z-score/IQR detection, AI explanation and human review before an anomalous action executes.
Evidence: Dashboard, credential-free demo mode, audit trail, documented architecture and 149 tests.
Outcome: The statistical model decides, AI explains and a person approves: separating those responsibilities makes the decision reproducible and auditable.
Scope: Its scope is a hackathon project on testnet. It has not undergone a security audit or production hardening, and its operating-loop state does not persist across restarts.

Code ↗Analysis ↗

Data products · Model selectionMaintained

LLM Radar

A daily dashboard bringing hosted and open-weight markets together to compare cost, context, adoption, licenses and restrictions.

Context: Choosing an LLM meant reconciling pricing pages, catalogs and adoption signals that change every week.
Design decision: An OpenRouter + Hugging Face → Python → Quarto → GitHub Actions and Pages pipeline, refreshed daily with no server.
Evidence: Live dashboard, pipeline code, methodology, searchable catalogs and comparable visualizations.
Outcome: License and gated access are governance decisions for self-hosting and commercial use; downloads alone lag when identifying new models.
Scope: The comparison is limited to catalog metadata; it does not evaluate response quality or benchmarks. Its accuracy depends on upstream public APIs.

Code ↗View live ↗

Observability · AI operationsMaintained

ccfuel

A local gauge separating the official quota percentage from its own attribution of non-cached work by project and session.

Context: The official gauge shows how much Claude Code quota was consumed, but not where the work concentrated; raw token totals are dominated by cache reads.
Design decision: It combines the official `/usage` gauge read through a PTY with a local proxy summing output, uncached input and cache creation by session.
Evidence: Local dashboard, privacy-safe synthetic demo fixture, methodology, API and regression tests.
Outcome: Quota percentage and activity proxy are different units: keeping them separate supports scheduling decisions without inventing a conversion Anthropic does not publish.
Scope: It requires authenticated Claude Code and local transcript access. The proxy is an attribution heuristic, not an official quota formula.

Code ↗Analysis ↗