Goodwill Agent
Voice-first tagging at the warehouse rack — my Master's capstone project at Carnegie Mellon, for Goodwill of Southwestern Pennsylvania.
Voice-first tagging at the warehouse rack — my Master's capstone project at Carnegie Mellon, for Goodwill of Southwestern Pennsylvania.
Researcher & Product Builder — I led the user research, product design, and engineering.
1. Huge impact: processors tagged +123% more items per hour in the pilot test.
2. Rare scope: I got to take Goodwill's first internal AI tool all the way from field research to working code.
Goodwill processors hand-tag ~10M donated items a year — every single garment gets tapped through a desktop app before its tag can print. And with ~50% annual turnover, pricing knowledge keeps walking out the door.
Small improvements are worth millions here: a $0.50 lift in average selling price models out to roughly $5M a year for the mission.
8 processor interviews, 8 think-aloud sessions, and 20 shopper interviews later, one insight stood out: the bottleneck is cognitive, not physical. Processors lose time guessing sizes, recalling brands, and walking to the tagging computer — not moving clothes.
They already invent their own decision aids — backwards hangers for up-pricing, batching high-value items. Our job was to formalize those aids, not overwrite them.
Three questions anchored everything we built afterward.
Three processors withdrew from our Wizard-of-Oz study the moment they learned it involved AI. The nine who took part overrode nearly half of the AI's suggestions — almost always pricing downward, to match what actually sells in their store.
8 of 9 told us the same thing: AI should be backup, not boss. Adoption — not accuracy — became the real design problem.
Four assumptions didn’t survive contact with the warehouse floor — and each pivot pushed us away from building another pricing algorithm, toward designing the human workflow that makes AI usable.
That reframing is what led to Goodwill Agent: not a smarter screen, but a way to keep the processor’s hands on garments and their judgment in charge.
We spent three weeks finding out. Four mics × three transcription tools in a noisy, echoey space: a clip-on mic hit 97–98% accuracy. Timed voice-vs-touch runs with 5 processors: voice was faster for most — and the one they preferred.
The survey said design for skeptics: of 131 Goodwill employees, only 8% trusted AI recommendations. And the existing platform's built-in voice needed rescuing more than half the time. So we built our own voice flow.
Try the voice prototype we tested →Goodwill Agent is a single Windows app — a slim, always-on-top voice strip docked beside Goodwill's existing Solutions tagging software. The processor keeps their hands on the garment and just says what it is.
The agent turns that sentence into a validated tag command — offline or cloud speech, fuzzy and phonetic matching onto an apparel vocabulary — then clicks through Solutions itself. The tag prints at the rack. No taps, no walking.
So trust was engineered in. The agent never touches the worker's Solutions credentials — it uses the session they signed into. Commands are idempotent and processed one at a time, through a verify-act-verify state machine that stops and asks for a human whenever the UI looks unexpected.
After two on-site visits with Goodwill IT, the agent ran the full tagging workflow on the live Solutions system (DGR.exe) in Goodwill's controlled test environment — and it ships as a self-updating install.
In the pilot test, processors tagged +123% more soft-line items per hour, with a 15% average lift in sell-through — Goodwill's first internal AI workflow tool, in real use.
The research traveled, too: Goodwill SWPA leadership distributed our readouts to regional Goodwill chapters across the country.