Visual Recommendation Feature: Scoping and Business Case for Fashion-Insta

Context. Fashion-Insta is a fashion retailer turning over 10.4 M€ a year, split evenly between its web channel and eleven physical stores in France, with 950,000 unique visitors a year and a catalogue of 840 references across nine categories. The company wanted a visual search feature on its mobile app: photograph an outfit, get the closest matching items from the catalogue. This was a scoping assignment, not a build. Nothing was implemented beyond a benchmark of the segmentation model, and the deliverable was a decision package presented to an executive committee in June 2026: scope, architecture, staffing, phasing, cost, return on investment and GDPR position. It is the project in this portfolio closest to the consulting side of the role I am targeting.
Problem. The company framed its own need as a data problem, product data existing but not being centralised, with a target of 400,000 active app users and a projected 14% lift on web revenue and 4% in store over 24 months. The real difficulty was different: an executive committee cannot approve a project on a technical intention. It needs to know what will be built first, what will be deliberately left out, what evidence will settle whether to continue, what it costs, when it pays back, and where the regulatory exposure sits. Producing that without writing the system is the entire exercise, and every number in it has to be traceable to a stated assumption.
Constraints. The audience set the form: a committee reads a business case, not an architecture. Photographs of people are personal data and some of them contain faces, so GDPR could not be a closing slide, it had to constrain the architecture. The catalogue is small, 840 references, which makes vector search trivial at current volume and means the design has to be sized for the target rather than the present. And the whole proposal had to remain arguable without an implementation to point at, so every technical choice needed a comparison and a stated reason.
Approach. I narrowed the proof of concept to one feature, visual recommendation from a photograph, and wrote down what was excluded before anything else: virtual try-on, colour and style transfer, preference or brand-based recommendation, account management. I then fixed three success criteria in business terms before any build, so the go or no-go decision could not be rationalised afterwards: segmentation accuracy measured in IoU, recommendation relevance measured as at least one item in five in the correct category, and catalogue coverage. For the technical approach I benchmarked SegFormer B3 for segmentation, reaching 0.69 mIoU, Fashion-CLIP for embeddings and Azure AI Search for vector retrieval, and compared this against a multimodal LLM alternative that would describe the photograph in text and search on that text. I rejected the LLM route on four grounds: it loses fine visual information in the text bottleneck, it costs more per token at scale, it has higher latency, and it sends photographs outside the EU. The architecture indexes the catalogue once, offline, and reuses the same Fashion-CLIP model at user inference so both sides live in the same vector space. I assigned development, deployment, maintenance and supervision for each of the seven bricks, with a deliberate split where data scientists build the ML components and MLOps operates them, and the Tech Lead supervises the two decision points of the flow, segmentation at the entry and vector search at the similarity decision. The plan runs in three phases, proof of concept over six weeks, MVP over ten, run-up to production over three, with a weekly project committee throughout and a steering committee closing each phase.
Result. The package answers the committee's questions with figures it can audit. Staffing costs 69,650 € one-shot, 21,350 € for the proof of concept, 35,560 € for the MVP and 12,740 € for the run phase, computed from stated day rates over 22 working days a month. Recurring cost after go-live is 3,718 € a month, of which only 110 € is cloud infrastructure: the running cost of this system is people, not Azure, which is the opposite of what a committee usually expects and it changes where the scrutiny should go. Against a modelled gain of 39,000 € a month the net is 35,282 €, cumulative gains overtake cumulative costs in the second month after go-live, and the 24-month surplus approaches 777,000 €. The sizing rests on explicit assumptions I stated rather than buried: roughly 7 MB per request, a vector index of 1.68 MB, and 1.2 million monthly requests derived from 400,000 users making three searches each. The GDPR section maps four risks to four architectural answers, and one of them is the point I would defend hardest: rejecting the multimodal LLM on precision, cost and latency also removed an article 46 transfer of photographs outside the EU. The privacy position and the performance position are the same decision, not a compromise between two. Face blurring happens on the device before any network call, and request logs carry a technical session identifier with no name or email.
Limits. This is a proposal, and the only thing in it that was measured is the 0.69 mIoU of the segmentation model. Everything downstream of that is a model of a system rather than an observation of one. The revenue gains come from the client's own projection, so the payback date inherits that assumption entirely, and I ran no sensitivity analysis and no downside scenario: if the gain figure is half of what is claimed, the business case still works but the conclusion I presented does not show it. The relevance criterion I set, one relevant item in five, is a low bar that I chose myself to make the go decision reachable, and it would need to rise before the feature reached users. Catalogue coverage was defined as a criterion but never given a threshold. And the plan assumes the client's data can be centralised without difficulty, which was the problem the company originally described and the one part of it that the scoping never priced.