Launch Your AI Startup: A Step-by-Step Guide
▼ Summary
– Julie Bornstein, a seasoned fashion e-commerce executive, founded the AI startup Daydream to help customers discover clothing, but found the implementation far more difficult than expected.
– Despite the proliferation of AI apps since ChatGPT, studies show they have not yet delivered a significant, widespread productivity boost, with many enterprise projects failing to provide measurable value.
– Daydream’s core challenge involves interpreting complex, nuanced customer requests (like event-specific outfits) and reliably matching them to products, a process hampered by AI inconsistencies and “hallucinations.”
– To address these issues, Daydream delayed its launch, hired a top technical team, and moved from using a single AI model to an ensemble of specialized models for attributes like color, fabric, and season.
– The startup must bridge the gap between “shopper vocabulary” (personal, contextual desires) and “merchant vocabulary” (product categories and attributes), using both language and visual models to understand customer needs.
Julie Bornstein believed launching her AI fashion startup would be straightforward. With an impressive background in digital commerce, including roles at Nordstrom and Stitch Fix, she felt perfectly equipped to build a company that uses artificial intelligence to help people find ideal clothing. The reality, however, proved far more challenging. Her experience with Daydream, a venture backed by $50 million from investors like Google Ventures, highlights a critical gap in the current AI landscape: transforming impressive technological demonstrations into genuinely useful, reliable applications is an immense hurdle. This difficulty helps explain why, despite a surge in AI apps, the promised revolution in productivity has yet to materialize on a broad scale.
Bornstein’s initial concept seemed clear: leverage AI to solve complex fashion dilemmas, connecting users with perfect garments and taking a commission. Many assume this simply involves plugging into a powerful language model’s API. The process, however, quickly revealed profound complexities. While securing partnerships with over 265 retailers offering millions of products was manageable, interpreting a simple request like “I need a dress for a wedding in Paris” became a labyrinthine task. The system must deduce context, the user’s role, the season, formality, and desired impression. Even with those parameters, inconsistency and “hallucinations” from AI models created major reliability issues. During extended beta testing, a user asking for a dress to make a “rectangle” body shape look like an “hourglass” might receive suggestions for dresses with geometric patterns, a literal and useless interpretation.
This forced a strategic pivot. Bornstein postponed the planned 2024 launch and strengthened her technical team by hiring Maria Belousova, former CTO of Grubhub. Belousova assembled a team of engineers attracted by the unique challenge. “Fashion is such a juicy space because it has taste and personalization and visual data,” Belousova notes, emphasizing the unsolved nature of the problem. Daydream’s mission involves solving it twice: first by accurately interpreting the customer’s often nuanced or emotional language, and second by mapping that understanding to the structured, attribute-based vocabulary of merchant catalogs. A query like “I need a revenge dress for a bat mitzvah” requires deep contextual comprehension.
To bridge this gap, Daydream moved beyond relying solely on language models. The company now employs an ensemble of specialized AI models, each handling a specific aspect like color, fabric, season, or location. This approach, combined with visual AI that analyzes products in a nuanced way, has improved results. Bornstein explains they found OpenAI’s models excel at understanding clothing from a world-knowledge perspective, while Google’s Gemini offers speed and precision for different tasks. The platform remains in a technically extended beta as it refines this complex system, a testament to the patience and persistence required to build truly functional AI applications that deliver measurable value.
(Source: Wired)
