Zach
Lichtman
Infotainment DevOps @ Hyundai · AI/ML @ WiscoHumanoids
I build the thing I wanted to exist, then make sure it actually works. Right now that means deciding whether a voice assistant is safe to ship to millions of cars.
Toolkit
Systems
Languages
AI / ML
AI Safety
Technical Fellow, Wisconsin AI Safety Initiative
Cybersecurity
Northwestern CTD
Frontend & Design
Backend & DevOps
Cloud & Infra
Spoken
Education
University of Wisconsin Madison
B.S. in Computer Science
Experience · 2026
Hyundai · HATCI
Software Engineering InternInfotainment DevOps & Production Software
HATCI is Hyundai Motor Group's R&D center in the U.S., where Hyundai, Kia, and Genesis engineer what ends up in the car. I work on the tooling that decides whether it is safe to ship.
VERDICT
Architected VERDICT, an evaluation harness deciding whether a production LLM voice assistant is safe to ship to millions of vehicles: it holds multi-turn spoken conversations with each assistant, grading them with 25 scorers across four channels (semantic, acoustic, interactional, stability) from deterministic assertions to Prometheus-2/G-Eval/FLASK judges. Adopted into Red Hat's Automotive Suite for use by other carmakers.
Built internal developer tooling on the Red Hat OpenShift platform standardizing software development for 1000+ developers across Hyundai Motor Group (Hyundai, Kia, Genesis): automated CI that deploys and validates builds on real hardware benches and device emulation, plus GitOps delivery pipelines.
Partnered with Amazon's voice team and Michigan State University researchers on production integration and pre-release validation of in-vehicle voice features.
Consulting · 2025-2026
Brainsprout & Raden Wellness
Turning MRI volumes into brain models a neurosurgeon can teach and testify from.
Engineered PyTorch/MONAI pipelines rendering anatomically accurate 3D brain models for neurosurgical education & expert testimony (Brainsprout, led by Dr. Michael Walsh, Head of Neurosurgery at Northwestern Medicine).
Guiding AI feature decisions and HIPAA-compliant data handling for Raden Wellness's health platform.
Open source
Problems I hit, and engineered away
PowderMeet
Where should two skiers separated across a mountain meet? No app could answer, so I solved it as a graph problem.
Live across 159 Epic and Ikon resorts
Born from a winter-break problem no app could answer: where should two skiers separated across a mountain meet? Invented the solution: Dijkstra over the resort's OSM trail/lift graph minimizing the slowest skier's arrival (max of the two) under trail topology, lift queues, and ability gaps, with α-weighted lift-wait penalties and edge-ID tiebreaks so both phones independently converge on the same meeting point; live across 159 Epic and Ikon resorts.
Server builds an identical graph manifest for every device (Overpass + DEM data, quantized to 15-minute buckets, 0.5°C, and 10 m elevation for consistent hashing); Geohash-6 Supabase Realtime channels scope presence fan-out by area; a Postgres state machine (CHECK/trigger guards, RLS scoped to accepted friendships) tracks meet-up requests.
Heartable
A Valentine's mixtape that refused to care which streaming service owned each track.
Built to make a Valentine's mixtape that could span Spotify, Apple Music, and other services: unifies them into one master library behind a provider-adapter layer, where a multi-source player plays each track through the service that owns it, skip-edit song versioning enables mixtape-style playback, and library snapshot/restore round-trips as idempotent CSV for cross-account migration. Also tracks listening stats and a friends leaderboard.
Ground-up SwiftUI rewrite of a React Native app (SPM-only, zero CocoaPods), reusing the existing Supabase backend unchanged: Postgres with RLS, RPCs, edge functions, and server-side OAuth token refresh per connected service.
Robotics
Teaching hands to move, and rovers to know where they are
AI/ML Engineer, AlohaMini
Teleoperation data collection and imitation-learning policy training (Hugging Face LeRobot, ACT) for the AlohaMini dexterous hand.
Contributed stale observation detection and logging to the LeKiwi client to surface dropped frames during data collection.
Software Engineer, GNSS Navigation
Validated GNSS navigation code for the club's autonomous rover, field-testing across varied terrain and signal conditions.
AI safety
Making models safe, and teaching people to
Safety Scholar
Conducts collaborative AI-safety research, backed by awarded model API credits and research resources (20+ papers published org-wide).
AI Enablement & Training
Authored the team's technical training materials on AI development and AI safety, then co-taught engineers to use the tooling properly: agentic loops, agents, and evaluation practice with safety in mind.
Instructor
Taught introductory Python to first-time programmers, guiding students from zero to shipping their own mini games.
Also
Other things worth knowing about
JRNL-iOS
Contributor on alichtman/jrnl-ios
Contributor to an encrypted, git-backed journaling iOS app. Shipped Time Machine, a per-entry commit-history browser that reconstructs entry state at any commit via libgit2 blob reads, and bridged Clibgit2 push/fetch/merge into reusable Swift Repository extensions.
Built the GitProviderForPAT protocol family, unifying GitHub, GitLab, and self-hosted PAT sign-in behind one Swift interface.
Independent Bug Bounty
Cybersecurity Researcher
Identified and responsibly disclosed an IDOR (insecure direct object reference) vulnerability in OpenAI's ChatGPT.
WhatIsTheMove
Anthropic · ClaudeHacks'26
Campus connectivity, solved with a spatial index and a model that writes the icebreaker.
SmartCards
UW CDIS · MadHacks'24
AI flashcard generator that extracts key terms from user-supplied text and auto-generates study cards with LLM-written definitions.
Claude Builder Club
Anthropic Student Developer
Build with the Claude SDK and help other students ship their first Claude-powered projects.
Wisconsin Business Review
Author, Technology Committee
Published “Rational Machines; Irrational Standards”: AI drivers already outperform humans yet are held to irrational safety standards.
