AI
The Curly Hair Market Gap
Black women spend 9 times more on hair products than other groups, yet the category still delivers poor results. Research the team gathered during R&D showed that 43% of Black women use five or more products simultaneously, cycling through shampoos, conditioners, and styling creams with no reliable signal about fit.
“43% of black women use 5 or more” — Alicia Adams
The root problem is a trial-and-error default with no feedback loop. Textured hair products cost more because of a minority hair tax, and dissatisfaction remains high anyway. HairMatch was built to replace guesswork with a hair profile tied to specific products.
Live Hair Scan Demo
Matt and Alicia ran the live hair scan on stage (03:40) on a real iPhone. Alicia lined up her hair in the camera, took two pictures under good lighting, and tapped “get your match.” The app sent both images to Azure AI Foundry, where a visual transformer processed them and returned a result in seconds: normal porosity, curly texture, type 4A.
Two images are required by design. A single shot is not enough for maximal accuracy, so the app instructs users on framing and lighting before the scan runs.
Recommendations and Ongoing Hair Tracking
The hair profile drives a curated product list filtered by hair type. Users browsing shampoos see only options matched to their profile, with attributes like sulfate-free and clean beauty standards surfaced from the R&D-sourced product database. A link out to purchase is one tap away.
A second feature, “HairMatch Learns,” surfaces Alicia’s accumulated hair knowledge as contextual guidance. Users can run daily scans to track whether products are actually improving their hair over time. Progress is visible, not assumed.
From GPT-4o Prototype to 95% Production Accuracy
GPT-4o Vision, accessed through Azure OpenAI, made the product possible without custom training. In August 2024, the team sourced 150 people with known hair types through Alicia’s network, collected photos, and tested off-the-shelf transformers. GPT-4o with a single prompt hit 90% accuracy immediately. Multi-shot prompting plus controlled lighting pushed it to 95%, and that configuration shipped to production.
Fine-tuning via Azure AI Foundry is in progress to lower costs further. Matt closed with a transfer principle: the same approach applies to any vision classification task, not just hair. Start with GPT-4o before assuming custom training is required.
Notable Quotes
43% of black women use 5 or more Alicia Adams · ▶ 2:49
Vision language models have been a revolution over the past. Matt Steele · ▶ 7:42
It’s about anything related to vision. Matt Steele · ▶ 10:01
Key Takeaways
- GPT-4o classified hair type at 90% accuracy from a single prompt with no custom training data.
- Controlled lighting and two-image input pushed accuracy to 95%, which was good enough for production launch.
- The same vision classification pattern applies to any image-based task that would have cost capital to build five years ago.