John Boen

Data science platform for optimizing a 13-image calendar from 12,000+ NASA Artemis II photos — framed as constrained portfolio optimization, not top-N ranking.

• Multi-modal feature store: CLIP vision embeddings (512-dim), sentence-transformer text embeddings, and perceptual metadata clustered via HDBSCAN with weighted fusion
• Beta-Binomial posterior scoring with exposure weighting, Elo ratings from pairwise votes, and Borda counts from category rankings — fused into a composite score with uncertainty penalties
• Hungarian algorithm for provably optimal image-to-month assignment with diversity constraints enforced via CLIP cosine similarity

Multi-provider exam study app supporting 10 certification vendors with a provider-agnostic quiz engine, zero production dependencies, and an Atlas design system.

• Provider-agnostic core: new vendors plug in by adding JSON data files and a branded HTML page — no shared code changes needed for the engine or scoring
• Stateless QuizEngine class with pure business logic, no DOM dependency — fully testable with idempotent answer submission and temporal hint tracking
• Token-first Atlas design system: all visual decisions flow from CSS custom properties with single-attribute dark mode toggle and zero client-side JavaScript

Regulatory compliance readiness platform covering 12 data regulation frameworks with self-assessment tools, control inventories, and advisory guidance.

• Hub-and-spoke architecture: hub site discovers topic spokes at build time by scanning sibling directories for site.json manifests validated against a Zod schema
• Zero client-side JavaScript — navigation uses CSS-only hamburger toggle, status badges are pure data-driven rendering, entire site is static HTML
• Content collections pattern with four data dimensions (controls, advisory, services, tools) per regulation, enabling type-safe Astro collection queries

Labor market analytics platform that ingests federal data (SOC, O*NET, BLS) into a four-layer DuckDB warehouse and serves occupation insights through a pre-rendered static site.

• Four-layer immutable warehouse (Raw → Staging → Core → Marts) with schema drift detection that blocks publication on unexpected column changes
• Static site served from GitHub Pages using a fetch() shim that intercepts API calls and redirects to pre-built JSON — full API ergonomics with zero server cost
• Multi-vintage time-series with CPI-deflated real wages, crosswalk-aware taxonomy evolution, and as-published vs. comparable-history modes

Knowledge management system that harvests markdown lessons from multiple GitHub repos into a searchable static site with a RAG chatbot backend.

• Adapter pattern abstracts LLM and vector operations across four cloud stacks (local Ollama/ChromaDB, AWS Bedrock, Azure OpenAI, GCP Vertex) with lazy SDK imports and mock-friendly testing
• Seven-rule gap detection identifies corpus weaknesses from retrieval scores, triggering automated GitHub discovery and candidate lesson staging
• H2-boundary chunking preserves semantic structure (789 chunks from 116 lessons) instead of token-window splitting

Prompt-driven software development lifecycle that converts product ideas into structured, phased implementation plans executable end-to-end by AI agents.

• Document-as-state model: all project status lives in markdown task tables with explicit state transitions (Open → Started → Completed/Blocked) — no external tooling required
• Self-contained phase plans generated by expand stage include schema, imports, and patterns so an implementer can execute holding one file and the codebase
• Deployed across 6+ projects (resume, data-readiness, SOC 2, GTM Leads) — reproducible and battle-tested