TL;DR: Ran a stress-test and system-architecture audit on a live D2C fashion AI shopping assistant, found the root failure was a stateless pipeline (persona signals weren't persisted into retrieval ranking, and checkout was loosely coupled to merchant auth state), then shipped a scoped MVP — a Fit Confidence Badge — as the trust-building wedge before higher-RICE features. Outcome labeled modeled/projected for impact; the MVP execution plan and sprint breakdown are fully specified.

Context & Strategic Framing

Verifast builds AI shopping assistants for D2C fashion brands — chatbots meant to understand products, read intent, recommend, handle objections, and close sales, not just answer FAQs. I ran a stress test against the seven core requirements (product intelligence, intent detection, smart recommendations, objection handling, lead timing, multilingual support) to find where the actual system breaks, rather than proposing net-new features on top of an unaudited base.

Problem — What the Stress Test Actually Found

What works: visual discovery (festive product carousels), a visible "ADD" button that supports impulse conversion.

Critical failures, in order of severity:

1. Persona hallucination — male users receiving women's dress recommendations. This is an immediate trust-loss event, not a minor relevance miss.
2. Trust gaps — no verified/in-stock signals shown on product cards.
3. Checkout friction — OAuth errors and missing size selection block purchases at the exact conversion moment.

Architecture root cause: the pipeline (Intent → Persona Tag → Retrieval → Templated Reply) is linear, but persona signals aren't persisted into the retrieval ranking layer — so a correctly detected "male user" tag doesn't actually constrain what gets retrieved. Separately, checkout is loosely coupled to merchant auth state, which is why auth errors surface as customer-facing failures instead of being handled gracefully upstream.

Six Unstated User Pain Points

Beyond the stated requirements, I surfaced: trust deficit (no provenance/verification signals), size uncertainty (missing guidance causes hesitation and wrong orders), hidden offers (promotions surfaced too late in the flow), checkout breaks (OAuth/validation errors lose completed carts), cognitive overload (navigation jumps break conversational flow), and decision paralysis (no quick reassurance signals for higher-risk purchases).

Competitive Landscape — Where to Differentiate vs. Match

Verifast's agentic, sales-closing approach differentiates it from support-only bots like Gorgias — but the strategic gap to close is Vue.ai's depth in fit/sizing intelligence, since that's the single largest driver of returns in fashion e-commerce.

Prioritization — RICE, and the Non-Obvious Sequencing Call

Strategic decision — built Fit Confidence Badge first despite the marginally lower RICE score: it builds the trust required for the Inline Size Selector to actually convert. Users won't commit to a quick in-chat purchase if they aren't confident about size — sequencing by dependency, not by raw score, is the senior-PM call here.

Explicitly rejected: 3D Try-On (heavy engineering, long ROI horizon) and a Mandatory Size Quiz (adds friction and drop-off before the user has any reason to trust the system yet).