Projects
I help talent teams improve process, optimize their TA stack, and build infrastructure that makes hiring swift and efficient. The goal is always the same: eliminate the rote work so recruiters can spend more time with candidates, where it actually matters.
Every service I offer is grounded in having built the same systems myself. The projects below are how I trial modern technologies before bringing them to clients at scale.
From Lever to Greenhouse, Greenhouse to Ashby, or building from scratch. Scoping, configuration, data migration, and integrations with HRIS, scheduling, and interview intelligence tools.
Finding where hiring accumulates drag and eliminating it. Workflow redesign, structured interview frameworks, SLA accountability, and the reporting that keeps it honest.
The right tools, configured correctly: AI scheduling, interview intelligence, and custom agentic workflows that remove manual work from your team's plate permanently.
An experienced head of talent for teams in transition. Full ownership of hiring strategy, team direction, stakeholder management, and executive reporting.
A structured evaluation of your current tooling: what's working, what's not, where the gaps are, and a concrete plan for what to change and in what order.
Dashboards that turn recruiting activity into business intelligence. Hiring velocity, stage conversion, DEI metrics, and the data structures that make future analysis possible.
Building AI agents that act across tools on a schedule, without being prompted. A self-hosted Hermes agent delivers a weekday morning brief to Telegram, built from the to-dos and goals in my notes, plus a weekly recap of the email that needs attention. Next up is an email triage agent that classifies incoming mail with a local model and escalates the uncertain calls to Claude. I use Obsidian as a thinking environment for all of these projects, capturing reasoning and decisions in a way that feeds directly back into the agents as context.
A direct proof of concept for the AI and automation work I do with clients: real agents, eliminating real manual work, running every day.
ActiveA self-hosted home lab: an Intel NUC running Debian with Plex and containerized services, and a Mac Mini running local models through Ollama. Sensitive work, like classifying my own email, stays on my own hardware, and the setup doubles as a playground for agentic experiments and automation workflows.
Running AI at the infrastructure level means I can evaluate and recommend tools from the inside, not just the vendor's sales deck.
ActiveA sourcing platform that ranks software engineers by their public work, analyzing GitHub contributions and community signals instead of résumés. Vamo also plugs into AI assistants like Claude and ChatGPT, so searches can run inside the tools a recruiter already uses. Evaluating how its rankings hold up against traditional sourcing for technical roles.
For technical hiring, what an engineer has actually built often says more than their résumé. Tools like this could change where sourcers start a search and how they build shortlists for hiring managers.
EvaluatingA multi-stage ETL pipeline built on a personal music library of 55,000 tracks. The extract stage ingests metadata from audio file tags and the Plex API. The transform stage normalizes genre taxonomies across 150+ canonical categories and extracts acoustic features (BPM, key, energy, danceability, timbral fingerprints) using Essentia, with a separate ML inference pass using a Discogs EffNet model for genre classification across 400 styles. The load stage writes to SQLite, powering a self-hosted playlist engine that queries the processed data by acoustic parameters and genre filters and pushes results directly to Plex.
The pipeline runs nightly via cron, is resume-safe across interruptions, and handles subprocess isolation for files that cause library crashes. The same dataset that started as a music project is now a foundation for broader ML experimentation. It also audits the library itself: cross-referencing every album against MusicBrainz flagged 955 incomplete albums and 215 duplicate or split entries for cleanup.
Tools: Python, Essentia, TensorFlow, SQLite, Flask, Plex API, MusicBrainz API, mutagen
The ETL architecture, ML pipeline, and data modeling here translate directly to recruiting analytics: stage conversion, time-in-stage anomalies, and the signals that predict outcomes. The library audit is the same exercise as cleaning up ATS data before a migration.