Delivery leader who still builds — data engineering and applied AI, translating enterprise and federal initiatives into systems that ship.
Delivery leader who still builds — data engineering and applied AI, translating enterprise and federal initiatives into systems that ship.
Lead the VA technology portfolio across delivery, strategy, and stakeholder alignment — translating enterprise IT and AI initiatives into funded, executable programs.
Managed the Infrastructure as a Managed Service (IaaMS) and Storage on Demand (SOD) programs, driving delivery, contract compliance, and client satisfaction across multi-year VA engagements.
Oversaw delivery of the Storage on Demand program, coordinating technical teams and federal stakeholders to meet SLA commitments and reporting obligations.
Supported the Enterprise Performance Measures Support (EPMS) program — capturing requirements, aligning technical outputs to KPIs, and bridging business and engineering teams.
Led procurement on the Data Center Acquisition T4 (DCAT4) program — developing competitive RFP structures, evaluating proposals, and translating technical requirements into acquisition strategy.
Administered and awarded competitive and sole-source contracts under FAR/DFARS, managing the full acquisition lifecycle from requirements through closeout.
A small fine-tuned retrieval model that maps VA National Drug File names to the exact RxNorm clinical drug — ingredient, strength and dose form together — and abstains when the string doesn't decide. 93.1% exact match on drugs whose ingredients it never saw in training, published with its dataset and two write-ups.
Python·PyTorch·sentence-transformers·DuckDB·Weights & Biases·Hugging Face
Published
A fine-tuned bi-encoder that maps VA drug names to RxNorm clinical drugs, with calibrated confidence so most answers can be accepted without review.
Published tools such as RxMap map VA drug strings only to the ingredient. This project asks how far a small trained model gets at the full clinical-drug level (SCD/SBD), where strength and dose form must also be right. RxNorm already links every VANDF name to its concept, so the labels are free. SapBERT fine-tuned with ingredient-matched hard negatives and a strength-normalizing preprocessor maps 93.1% of test strings (95% CI 91.8–94.1) to the exact clinical drug, against 50.9% for TF-IDF, with the right answer in the top five 98.4% of the time. Re-drawing the ingredient split puts the figure at 0.907 ± 0.019, and every later change is judged against that band rather than one favourable split.
Python · PyTorch · sentence-transformers (SapBERT) · DuckDB · scikit-learn · Weights & Biases · Hugging Face · Colab
A sold-comps and valuation data platform for micro-acquisitions — real sold prices and multiples for small online businesses, collected across marketplaces. Built and operated end to end by AI agents, with an owner who holds the payment accounts.
TypeScript·Cloudflare Workers·D1·Astro·Stripe·x402·Claude API
Live
Real sold prices and multiples for small online businesses — SaaS, content, e-commerce, newsletters, apps and domains — run by scheduled AI agents.
Buyers and sellers of small online businesses rarely see what comparable deals actually sold for. ExitComps collects sold listings continuously across marketplaces and turns them into a searchable comps database, a free valuation calculator, and a weekly “What Sold” digest. Revenue comes from Pro subscriptions and one-off valuation reports, paid by card through Stripe or in USDC over x402, which lets AI agents buy reports directly. Scheduled Claude Code sessions handle collection, marketing and bookkeeping through a scoped admin API; the owner holds the payment accounts and approves guardrail changes.
TypeScript · Hono · Cloudflare Workers, D1, KV & Workflows · Astro · Stripe · x402 (USDC on Base) · Claude API · Vitest · GitHub Actions
Autonomous trend-following bot trading $100 of real money on Kraken. Deterministic daily signals across BTC, ETH and SOL, a 40% trailing kill switch that liquidates and halts, and a public equity curve that regenerates itself daily.
Python·ccxt·Kraken API·pytest·cron·nginx
Live
Trend-following bot trading $100 of real money on Kraken, with a hard drawdown kill switch and a public equity curve.
cryptotrade trades BTC, ETH and SOL on Kraken through ccxt, checking hourly but deciding at most once a day per asset on closed candles. It holds an asset only while price sits above its 200-day average and the 20-day EMA leads the 50-day; otherwise that sleeve sits in cash. The interesting part is the risk plumbing rather than the strategy — live trading is gated behind two manual confirmations, a 40% trailing kill switch liquidates and halts, and every fill lands in a ledger that is never rewritten.
Python · ccxt · Kraken API · pytest · cron · nginx · rough.js
An autonomous AI-run data business — agents build the datasets, an agent operates the store, and the customers are other AI agents paying per request in crypto micropayments.
Cloudflare Workers·x402 Protocol·USDC/Base·GitHub Actions·JavaScript·REST APIs
Live
An autonomous AI-run data business — agents build the datasets, an agent operates the store, and the customers are other AI agents paying per request in crypto micropayments.
Datastand is a micro-business built to test whether AI agents can be paying customers of other AI agents. It sells small, machine-readable datasets — an npm vulnerability digest, a dev-signals newsletter feed — priced at a couple of cents per fetch, over the x402 protocol (HTTP 402 + USDC micropayments on Base, originally from Coinbase). A buyer agent hits the API, gets a 402 with payment terms, retries with a signed payment header, and receives the data plus a settlement proof — no accounts, invoices, or card rails. Humans can buy too, by sending USDC and redeeming the transaction hash on the storefront.
The distinguishing idea is that the business runs itself. A daily GitHub Actions pipeline rebuilds every dataset each morning and auto-deploys the gateway. A scheduled operator agent handles health checks, data-quality review, revenue ledgering, and one growth action per day, writing a plain-text report the human owner can skim. The system is architected to minimize the money it can lose: it can accept USDC (public address only) but holds no private key anywhere in the repo, CI, or server, so a full compromise of the code or infrastructure yields at most a handful of $0.02 datasets given away for free — never funds.
Cloudflare Workers · JavaScript (zero-dependency) · GitHub Actions · x402 protocol · USDC on Base · Coinbase x402 facilitator
Fork of Clawdmeter by Hermann Bjorgvin — native Windows daemon merged upstream in PR #57.
Native Windows host daemon for the ESP32 usage monitor — system-tray app with live status icon, login autostart, and auto-reconnect over BLE (WinRT). Strictly additive (no firmware changes); WSL-independent with a no-WSL-paths test guard.
Python·pystray·httpx·WinRT BLE·pytest·PowerShell
Shipped · 2026
Native Windows companion app for an ESP32 usage monitor — fork extended with system-tray, autostart, and WinRT BLE, merged upstream.
This project is an extension of Clawdmeter by Hermann Bjorgvin — not original work. The Windows host daemon was contributed back to the upstream project and merged in PR #57. Full credit to Hermann Bjorgvin for the original Clawdmeter concept and codebase.
Clawdmeter is an ESP32-based USB power usage monitor with a host-side app for live status. The upstream project targeted macOS/Linux; this fork adds a native Windows host daemon that runs in the system tray, launches at login, and reconnects automatically over BLE using the Windows Runtime (WinRT) Bluetooth stack — no WSL dependency. The daemon communicates over HTTP with the ESP32 firmware unchanged.
Python · pystray · httpx · WinRT BLE · pytest · PowerShell
Self-hosted middleware that syncs Kobo reading progress to Hardcover. Sits between a Kobo e-reader and Calibre-Web Automated, intercepting WiFi sync traffic and pushing progress automatically — also fixes generic book-cover display in the Kobo store. One-way sync with log-based monitoring.
Python·Docker·Docker Compose·Hardcover API·REST
Shipped · 2026
Self-hosted middleware that intercepts Kobo WiFi sync traffic and pushes reading progress to Hardcover via Calibre-Web Automated.
Calibre-Web Automated (CWA) ships its own Kobo sync, but it is unreliable for pushing reading progress onward to Hardcover. InkSync fixes that by sitting between your Kobo and CWA: it intercepts the device's WiFi sync traffic, forwards it to CWA unchanged, and pushes your reading progress to Hardcover automatically. As a bonus it corrects the generic placeholder covers that otherwise appear in the Kobo store. It runs as a containerized service alongside an existing CWA instance.
Python · Docker · Docker Compose · Hardcover API · REST
Strava activity name generator that replaces boring defaults with AI-generated witty quotes — 50+ activity types mapped to contextually relevant lines. Only renames generic activity names; encrypted token storage, rate limiting, and OAuth integration.
Python Flask·OpenAI API·Strava API·OAuth
Live
AI-powered Strava activity name generator — replaces generic defaults with contextually witty quotes across 50+ activity types.
SpicyFerret connects to your Strava account via OAuth and automatically renames newly uploaded activities that have a generic, app-generated title. Rather than "Morning Run" or "Afternoon Ride", it calls the OpenAI API to produce a short, contextually relevant, and occasionally witty replacement. The service only touches activities with default names — personal titles are never overwritten.
Python Flask · OpenAI API · Strava API · OAuth
Synthetic patient simulator generating clinically coherent, terminology-rich datasets — linked demographics, encounters, conditions, medications, labs, procedures, and care plans. YAML scenario/module-driven with ICD-10/SNOMED/RxNorm/LOINC/CVX coding and FHIR R4 output.
Python·Polars·DuckDB·Faker·PyYAML·pytest
In Development
YAML-driven synthetic patient simulator producing clinically coded, FHIR R4-compliant datasets for testing and research pipelines.
Synthetic Health generates realistic, clinically coherent patient records at scale — demographics, encounters, conditions, medications, labs, procedures, and care plans — all linked and coded against ICD-10, SNOMED CT, RxNorm, LOINC, and CVX. Scenarios and clinical modules are defined in YAML, making it easy to produce targeted cohorts (e.g., diabetic patients with comorbidities) or general population distributions. Output is FHIR R4, consumable by any standard health data pipeline.
Python · Polars · DuckDB · Faker · PyYAML · pytest
The site you're looking at — a framework-free single-page portfolio built from the ground up on a custom design-token system. Scroll-driven section reveals, an accessible project-detail drawer with focus trapping, and a privacy-friendly analytics layer. No framework, no bundler, no build step.
HTML5·CSS Design Tokens·Vanilla JS·GSAP·Jest
Live
A framework-free single-page portfolio built on a custom design-token system, with scroll-driven reveals, an accessible detail drawer, and a property-based test suite.
This site is a single-page portfolio rebuilt from the ground up — no UI framework, no bundler, no build step. Layout and components are styled with plain CSS driven by a custom design-token system, and behavior is handled by small, focused vanilla-JS modules. Scroll-driven animation uses GSAP with ScrollTrigger, with a graceful fallback when the CDN is unavailable so the site stays fully usable. Accessibility was a first-class concern throughout: focus management, a focus-trapped detail drawer, ARIA, and full keyboard navigation.
HTML5 · CSS Design Tokens · Vanilla JS · GSAP · Jest · fast-check
Have a project or role in mind? Send a note.
Thanks — I'll be in touch within a couple of days.