Mad Street Den found the mess behind the dress

A dress is an easy thing to recognise and a surprisingly difficult thing to describe. Is it blue or navy? A party dress or an evening dress? Does the neckline count as a scoop? To a shopper, these distinctions can feel trivial. To a retailer with millions of listings, they become the difference between a product that appears in search and one that spends its digital life waiting politely in the wrong cupboard.

This is a useful place to begin with Mad Street Den. Artificial intelligence has an unfortunate talent for making the ordinary sound theatrical. Here, the ordinary is the point. Someone must tell the computer what is in the photograph, reconcile that description with a supplier’s record, and make the result useful to the next person in the chain. The company built much of its early business around that work.

The dictionary hidden in a dress

Vue.ai’s product-tagging software uses computer vision and language processing to extract attributes from images and text. A business can map those attributes into its own taxonomy - the agreed vocabulary and hierarchy used to organise products. Once those records become more consistent, search filters and recommendations have something sturdier to stand on.

Consider an unnamed North American multi-brand retailer in a published Vue.ai case study. Different suppliers submitted different descriptions. Manual checks introduced mistakes of their own. Vue.ai says its system predicted tags for more than three million products over three months, added 69 new attributes and corrected 51.82% of the tags it processed in the live catalogue. These are vendor-reported results for that project, rather than a promise for the next customer.

The telling detail is the correction. An ecommerce catalogue can contain plenty of information and still be unreliable. Adding more fields does little good if the fields disagree. In this project, the system also used confidence values and feedback from quality checks. The glamorous shopping experience depended on a fairly sober exercise in deciding which descriptions deserved belief.

A designer, a neuroscientist, and a very large catalogue

Mad Street Den’s origins go back to 2013; Vue.ai arrived in 2016. The distinction matters because the company spent years working on the underlying technology before giving it a retail identity. Ashwini Asokan brought design experience from Intel. Anand Chandrasekaran brought a background in neuroscience. The official founding-team page also names chief scientist Costa Colbert.

Retail gave the technology a practical examination. In its 2019 Series B announcement, Vue.ai named thredUP, Tata and Macy’s among the businesses it had worked with. These customers had different commercial models, but each needed to connect product information with what a person might want to buy.

Resale makes that connection especially awkward. A conventional shop might have a run of identical shirts. A resale shop can have a vast collection of individual garments, each with its own brief moment in stock. When one disappears, the recommendation system must find another plausible choice.

“We have millions of items on our site and some customers don’t want to go through all of them.”

Chris Homer, thredUP CTO · Vue.ai WWD event report

In a historical account of thredUP’s GoodyBox service, Homer described customers completing style profiles and supplying Pinterest moodboards. Vue.ai worked alongside human stylists to help curate boxes. The arrangement is revealing: software organised possibilities; people remained part of deciding what should reach the customer. The ambition was personalised attention at a volume that would make an unaided stylist wince.

The machine must learn when it is wrong

At MercadoLibre, the problem was less about taste than admission. Marketplace images had to comply with rules about backgrounds, borders, watermarks and text. A manual review process had trouble keeping up. Vue.ai’s moderation software checked incoming imagery against those guidelines and incorporated feedback into retraining.

The case study reports a 70% fall in rollback rates over three months and a 51% reduction in seller contacts. A rollback here meant undoing an incorrect image decision. Those measures are more interesting than a decorative accuracy score. A wrongly rejected image can delay a listing and send an irritated seller to customer support. The cost travels beyond the model.

This suggests a sensible way to judge automation: follow the mistake. Who receives it? Who repairs it? What does it delay? A system that finishes its own task quickly can still make the surrounding business slower. Feedback and human review give the workflow a chance to improve rather than merely repeat itself at speed.

A wardrobe is only one kind of paperwork

The next move followed the work beyond product discovery. In November 2022, Mad Street Den acquired INTURN, software for managing slow-moving and excess inventory. Terms were undisclosed. Better recommendations might help a retailer sell stock, but they do not settle every question about what to do with merchandise that remains stubbornly unsold. INTURN extended the offer into inventory workflows and margin recovery.

Meanwhile, the underlying platform was finding other uses. Asokan announced Blox.ai in 2021 as a way to make the infrastructure behind Vue.ai available beyond retail. By late 2023, she described Vue.ai’s evolution into AI orchestration, pointing to repeated customer difficulties with fragmented systems, data maintenance and the cost of adoption.

Today Vue.ai markets data cleanup, product tagging, personalisation, intelligent document processing and workflow automation. Its intended buyers include operations, product, technical and business teams. Retail remains a clear source of expertise, while financial services, insurance, logistics and healthcare appear in its broader market offering.

FedEx supplied a concrete bridge into logistics. In May 2023, the company announced Mad Street Den as the first investment of its new FedEx Innovation Lab. The programme combined capital with commercial collaboration and access to FedEx’s network. For Mad Street Den, the opportunity was to apply its data and automation machinery to another business full of documents, exceptions and time-sensitive decisions.

The competitive claim is integration: connecting data preparation and models to business applications in one platform. A buyer can also assemble specialist tools or build internally. Larger data and enterprise AI platforms overlap with parts of the proposition. The useful comparison is therefore the actual workflow, its integration burden and its operating cost. A claim of broad coverage does not, by itself, establish superiority.

The capital raised and the price reported

Mad Street Den raised a $1.5 million seed round in 2015. A $17 million Series B followed in 2019. In January 2023 it announced a $30 million Series C led by Avatar Growth Capital, alongside existing investors Sequoia Capital and Alpha Wave Global. FedEx’s investment amount was undisclosed.

Then came a more uncomfortable number. March 2025 reporting described a cash-and-stock transaction with M2P Fintech worth roughly $10–15 million. YourStory reported an asset purchase covering intellectual property, contracts and employees, rather than a conventional purchase of the corporate entity. KeyValue, a product-engineering collaborator, subsequently said Mad Street Den had joined M2P. The company websites now carry M2P branding.

The reported asset price is not a verified valuation, and it cannot be treated as a tidy calculation of investor returns. Even so, it sits awkwardly beside the capital raised. Useful software, recognisable customers and venture financing do not automatically produce a lucrative exit. The public record supports that distinction without requiring an invented explanation for the sale.

Start with the queue

A historical founder interview offers a smaller, copyable lesson. Asokan put personal spending on the initial platform at about $100,000. She described annual enterprise contracts and said churn emerged when the company also pursued very small fashion startups. Serving both groups early proved difficult. The lesson was customer fit: identical features can belong to very different sales and support businesses.

For a prospective buyer, the practical starting point is a queue with a measurable cost: catalogue records awaiting checks, images awaiting approval, documents awaiting extraction. Establish the current error rate and review effort. Agree on the rules. Then test whether automation improves the whole process, including the work required to correct it.

Those are editorial implications of the cases, rather than a guaranteed deployment recipe. They become less attractive when volumes are small, the input is unusable, rules keep changing without an owner, or mistakes demand expensive review. Vue.ai’s advertised first-return-within-90-days framework is a vendor promise to examine against those conditions.

There is another habit worth borrowing. In a 2021 public post, Mad Street Den described women as about half its workforce, including leaders in engineering, analytics and product. That is a historical disclosure, but a specific one. The company’s interesting decisions often involved who got to participate: business teams defining categories, stylists reviewing choices, sellers providing feedback, and a broader group helping build the technology.

The dress at the beginning of this story has not become any more complicated. The organisation around it has. Mad Street Den’s useful insight was to look closely at the descriptions, decisions and handoffs that stand between seeing a thing and doing something with it. For anyone trying to put AI to work, that remains a good place to look.