A self-initiated, 11-day product discovery focused on the operational engine behind MPB's growth — Logistics, Warehouse Management, and the warehouse Operative app. It names the hidden "Listing Leg" no SLA measures, frames four strategic bets against a single throughput constraint, and backs them with three working AI-agent prototypes built end-to-end on a modern AI tool stack (n8n + Claude + Bolt.new).
About this project. A practice piece I built to demonstrate a complete, AI-powered product-discovery process end-to-end — from North Star to shipped prototypes — using MPB as a real-world subject analysed purely from public information. It is an independent analysis, not affiliated with or endorsed by MPB, and all figures are modeled estimates rather than audited results.
MPB is the largest global platform for used photography and videography equipment — a vertically-integrated managed marketplace that buys, inspects, grades, photographs, warranties and resells used kit in-house. It recirculated 615,000 items in FY25 (+9% YoY) against a used-camera market growing ~5.3% — taking share at roughly 3.7 points a year, with five market tailwinds firing at once.
That framing carries the whole thesis: demand is not the constraint — the warehouse is. In a Model-A managed marketplace, scaling recirculation means scaling operations. The single highest-leverage bottleneck — the Listing Leg between "Specialist-graded" and "live-on-site" — is not measured publicly, sits in no SLA, and is owned by no PM today. Closing it by even two days is modeled at a 5–8% lift to the North Star per quarter, with zero acquisition spend. This case study is the discovery that made the constraint visible and the three AI agents that prove it can be shipped.
The one-line mandate
Make the throughput constraint visible, then pull the gearbox conversion lever that turns demand-tailwind into recirculated items. Naming the constraint is the first step to managing it.
Eleven days of modular discovery, each step producing a named artifact — North Star → competition → market → process maps → FullStory simulation → funnel analysis → problem framing → OST & prioritisation → hypothesis workshop → business cases → prototypes. Every cost and impact estimate carries a confidence flag (🟢 high / 🟡 medium / 🔴 needs Day-1 data), and every 🔴 is logged as a measurement priority rather than hidden.
MPB's own published metric — two-sided, mission-aligned, and a leading indicator of margin. The PM-Logistics sub-NSM is Kit Throughput: items completing intake → live → first purchase, subject to a return-rate guardrail.
The hypothesis workshop was run with six independent agents — CPO, Head of Ops, Head of Data, Head of CS, Senior Specialist, Eng Lead — each proposing hypotheses blind. Convergence is therefore genuine agreement, not author bias.
A full trade-in on mpb.com (DE) surfaced the "PAID — done" seller view versus the warehouse view — and six concrete UX defects invisible in any public material.
A conversation with a Lead PM at Vestiaire Collective (adjacent luxury recommerce) validated several framings and productively challenged one — kept honest against a real operator's view.
The seller journey ends at "PAID — transaction closed." Everything that actually converts a bought item into a recirculation event happens after that, invisible to sellers and to every published SLA. Three compounding drains sit inside that blind spot.
The gap between graded and live-on-site. Conversion here is ~95–97% — the bottleneck is time, not conversion, concentrated in the manual pricing-approval path and an Operative-App-to-WMS handover, not in photography or (templated) copy.
A returned item is a second full Listing Leg — re-grade, re-photograph, re-list (6–19 days each). At industry-typical 5–10% return rates, that is an estimated 200K–800K item-days of trapped inventory a year MPB likely doesn't measure today.
Every misgrade feeds both the reject path and the returns path — the single most expensive quality leak. Reducing it 15–25% is modeled at ~€500K–€1M in annual savings, before any NSM effect.
Thirteen problem statements were scored on five dimensions (NSM impact, evidence confidence, PM direct ownership, stakeholder alignment, effort) and resolved into four strategic bets — each addressing a different shape of the throughput problem, so they compound rather than overlap.
| Bet | What it fixes | Year-1 spend | Likely impact |
|---|---|---|---|
| 🥇 Misgrading (P3.1) | Cost-loop quality — reject + return reduction | Process + rubric | 15–25% reduction; ~€500K–€1M/yr |
| 🥈 Kit-idle-time (P1.1) | Capacity engine — compress the Listing Leg | Workflow + thresholds | 5–8% NSM / quarter |
| 🥉 Ragehooks (P5) | Organisational speed — insight-to-action latency | €1–2K + n8n | Meta — accelerates every other bet |
| 🏅 AI-assisted intake (P2.1) | Capability foundation — remove the per-Specialist ceiling | €40–80K | 3–7% NSM / quarter (Year-1 ramp) |
Combined Year-1 investment ~€70–130K · combined likely NSM impact 10–15% per quarter additive by year-end · ~€1.5–3M annual margin add at steady state.
Each strategic bet is the lean-LLM business case made real — built end-to-end on a modern, accessible AI stack (n8n + Claude + Bolt.new), deliberately spanning three different LLM patterns to demonstrate range: text enrichment, structured-output decisioning, and multimodal vision.
n8n · Claude · Jira
A FullStory frustration event fires a webhook; Claude enriches it with a summary, priority and recommended action; a fully-contextualised Jira ticket appears within seconds. Insight-to-action latency drops from days to minutes — the unanimous workshop vote, and 3–5 dev-days in production.
n8n · Claude structured output
The pricing manual-review queue is the biggest Listing-Leg sub-component. This agent scores each item approve / review / decline against a rubric with few-shot examples, returning reasoning and confidence — widening the auto-approve band from ~60% rules-based toward ~85% AI-augmented, all in shadow mode.
Bolt.new · Claude Vision · n8n
Claude Vision inspects item photos for defects and compares against the Specialist's grade — but only after submission, surfacing a quiet confirmation on agreement or an amber "here's what I saw" panel on disagreement. Locks the workshop's design principle: AI as co-pilot, never autograder.
Why they compound
Ragehooks accelerates iteration on the other two; the Grading Co-pilot generates the labelled data that tunes Smart Pricing; Smart Pricing widens the Listing-Leg band that P1.1 depends on. Three agents, one loop — each with shadow-mode validation, so the worst case is "sunk cost, no operational harm."
The original business cases assumed traditional Eng + ML team builds — a combined ~€1–1.4M in Year 1. Reframing around foundation LLMs with prompting (no model training; the LLM is the model) cut that to ~€70–130K for equal expected NSM impact — a ~90% reduction. It ships in weeks not months, carries far lower commit risk, stays A/B-validatable, and is PM-led rather than Eng-bottlenecked. The upgrade to custom ML only triggers if vision accuracy plateaus or API costs exceed on-prem TCO.
| Business case | Investment | Payback |
|---|---|---|
| Ragehooks (H-5.1) | ~€1–2K + €25/mo | < 1 month · 🟢 low risk |
| Smart Pricing AI (H-2.1.3) | ~€30–50K Year 1 | 1–3 months · 🟡 medium |
| AI Grading Co-pilot (H-3.1.2) | ~€40–80K Year 1 | 4–8 months · 🟡 capability bet |
Most product cases pitch the what. This one pairs a defensible operating thesis — MPB's growth is constrained by warehouse throughput, not demand — with three working AI agents that demonstrate the how, built on a modern, accessible AI stack. It is honest about its unknowns: every low-confidence estimate is a documented measurement priority, not a hidden assumption. The discovery makes the constraint visible; the agents show it can be closed in weeks, not quarters.
Topics covered
Translating macro global events into SKU-level margin decisions in minutes. 3–7% margin recovery, 50+ data points automated, time-to-action cut from days to minutes.
Marketplace · B2CBootstrapped Halal meat delivery for 500K underserved Berlin consumers during COVID-19. 250K€ GMV in year 1, 151% customer growth.
Let's collaborate on your next high-impact product. I bring execution-focused discovery and AI-augmented delivery to complex operational problems.