Product Surfaces
Next.js, React, TypeScript, Tailwind CSS, shadcn-style UI
I use the modern web stack to turn workflow ideas into usable interfaces: dashboards, review queues, edit surfaces, and case-study pages that can be tested by a real user.
AI-assisted tools, grant searches, and practical systems
Andrew Cleary / 3R Programs
I work in banking compliance and have a strong interest in systems, new technology, and personal and professional growth. 3R Programs is where I document AI-assisted tools, workflow demos, and personal systems shaped by compliance experience, curiosity about how systems fit together, and a practical preference for tools people can actually use.
15+ years of experience
banking compliance and operations
SME
judgment in the workflow
AI Apprenticeship
learning AI by building live tools with AI
Work
Projects that turn workflow friction into working software.
The work falls into four areas: a personal workout tool that started the pattern, regulated-workflow prototypes, non-profit grant-search tools, and AI review systems that make recommendations reviewable, decisions explicit, and evidence easier to recover.
Personal Systems
The workout tool is where the pattern started.
It was not a banking project, and that is why it belongs here. It showed the broader working style: notice a real problem, build a narrow tool for the actual use case, troubleshoot what breaks, and improve the system until it fits the way it will really be used.
Compliance Systems
BSA/AML and Horizon Scanning apply banking compliance experience to focused software builds.
These are the projects closest to my professional background. They test how compliance work changes AI product decisions: where a model helps, where a person reviews, and what evidence has to survive for the output to be trusted.
Proof of concept
BSA/AML audit management
A narrow audit-management slice with findings, risk ratings, state transitions, role gates, and governed AI drafting. It is built to show domain-fluent product judgment, not broad feature coverage.
View case studyCurrent focus
Horizon Scanning
My current focus: a regulatory-change workflow for US banks, with source-grounded summaries, AI-proposed ratings, and human review.
View explanationNon-Profit Tools
The grant-search tools apply the same build discipline to organizations I support.
The useful pattern is intentionally simple: scheduled Claude searches, deduplication, Google Sheets updates, and a feedback loop from team decisions. PCH and NY3C use the same code pattern today; the shared engine is now in progress, with organization-specific profiles, credentials, sheets, feedback, decisions, and logs kept separate.
AI Review Systems
AI review systems make AI-assisted work more inspectable.
A central part of the way I work is building review into the process: tests, docs, handoffs, adversarial agents, the OpenAI Review Suite, and explicit verification before code or claims are treated as done. The goal is not confidence; it is work that can survive review.
Claude Code plugin
Engineering review system
A personal Claude engineering plugin with review, QA, security, accessibility, AI-output-integrity, and deployment-support skills. Hooks and review gates turn recurring discipline into workflow instead of memory.
View explanationAI reviewing AI
Adversarial agents
The agent pattern reflects a review discipline: one AI builds or drafts, another challenges the output for unsupported claims, weak code, overbuilt architecture, or security risk before it is treated as finished.
View explanationIndependent reviewer
OpenAI Review Suite
I built a Codex-native engineering review system with blind-first findings, evidence-backed reports, a GitHub exchange between independent reviewers, remediation, and verification. It is in active use and improves through each real project review.
View explanationTechnology Stack
Modern tools, used through a compliance-control lens.
The stack is not a standalone credential list. It shows how I scope and direct AI-assisted builds, understand the tradeoffs, and shape the control environment around the workflow.
Data, Controls, and Auditability
Postgres, Neon, Supabase, Drizzle ORM, RLS, pgvector
The database layer carries much of the control story: scoped roles, row-level security, audit logs, prompt records, embeddings, and separate rows for AI proposals and human decisions.
AI and Automation
Claude API, Claude CLI, Claude Code, OpenAI Review Suite, OpenAI embeddings, Python agents
I use AI where judgment or synthesis is useful, then pair it with structured outputs, prompt/version logging, review gates, second-model review, and simpler deterministic code where a model is unnecessary.
Verification and Delivery
Vercel, Clerk, GitHub Actions, Vitest, Playwright, pgTAP
The stack includes the operating discipline around the build: authentication, local-to-preview-to-production deployment, browser tests, database tests, CI checks, and review tooling.
Evidence-Led Work
Built so the reasoning is visible.
The projects are intentionally narrow. The credibility comes from the choices around each slice: what was scoped out, where human review sits, what gets logged, and how the implemented path is checked.
Practical scoping decisions
The work favors usable workflow choices over showpieces: management-response drafting over citation lookup, and Python plus Google Sheets over a heavier nonprofit web app.
Human decisions stay visible
AI output is treated as a proposal, not the final answer. Drafts, ratings, and matches are reviewed, accepted, dismissed, confirmed, or overridden before the result carries weight.
Evidence is part of the design
Prompt records, audit logs, source quotes, role gates, RLS policies, and review artifacts make the work inspectable. The practical question is always: what would a reviewer need to see?
Bounded scope with real checks
The projects do not claim full product maturity. They claim serious construction inside the slice that exists: tests, docs, validation notes, review passes, and clear limits.