Case types: Marketplace + Trust & safety (capstone)

1MARKETPLACE DIAGNOSIS FRAME23  1) Which side is cold? (here: demand)4  2) Is "hot" side real? (active pros last 30d vs registered)5  3) Liquidity metric: successful jobs / search or request6  4) Leakage: users find pro then go off-platform7  5) Unit economics by side: CAC, take rate, subsidies, refunds8  6) Trust: cancellation, no-show, quality complaints9  7) Category/geo focus: density before breadth1011  Data asks:12  · Search → contact → book → complete funnel13  · Time-to-first-book for new consumers14  · Pros active vs listed; response SLA15  · Off-platform leakage signals (phone reveals, chat export)16  · NPS / complaint reasons on failed jobs

1PLAN (DEMAND-COLD, SUPPLY-HEAVY)23  NON-GOAL: national expansion; more pro acquisition bonuses45  Bet 1 — Instant book SKUs in one category/geo6    Fixed-price cleaning packages with vetted pros, on-platform payment7    → removes quote friction89  Bet 2 — Trust package10    On-time guarantee, easy rebook, identity-verified badge that means11    something (revoke on no-show)1213  Bet 3 — Demand growth only in dense micro-geos14    Performance marketing + partnerships (property managers) where pro15    density already supports SLA1617  Bet 4 — Leakage reduction18    Masked comms + payment milestone release; value-add insurance1920  DEFER: cutting take rate as first move; it can race to bottom without21  fixing trust/friction.2223  METRICS:24    North star: completed jobs / weekly active demanders in focus geos25    Drivers: search→book; pro response SLA; cancel/no-show rate26    Guardrails: contribution margin after subsidies; refund rate27    Supply health: pro utilization (too low → churn of good pros later)28  Kill: if instant-book margin negative after 8 weeks without retention lift

1WEAK: 'Improve both sides equally: more ads for customers, more bonuses for2pros, lower fees, add AI matching, build an app redesign.'34STRONG: cold-side diagnosis · real vs zombie supply · trust/price friction ·5density before breadth · unit economics · leakage · utilization guardrail.

1LAYERED CONTROLS (defense in depth)23  L0 Prevention — account & entity4    · Employer identity verification tiers (email domain, company registry,5      payment instrument, manual KYC for high-risk)6    · Rate limits on new employer posting velocity78  L1 Detection — content & graph9    · Classifiers: phishing language, off-platform payment asks, "WhatsApp only"10    · Graph: reuse of job text, device/IP clusters, brand impersonation11    · Seeker reports with structured reasons1213  L2 Friction — progressive14    · Unverified employers: limited distribution, no resume PII unlock15    · High-risk postings: hold for review before apply1617  L3 Enforcement18    · Takedown, account ban, device ban, legal/escalation path19    · Kill switch: category-level pause if attack wave2021  L4 Recovery22    · Seeker alerts if they interacted with later-confirmed scam23    · Appeal flow for false employer enforcement with SLA2425  NON-GOAL: 100% manual review of every job (doesn't scale; destroys SLAs)

1METRICS + APPEALS + LAUNCH23  Outcome: scam-job contact rate per 1k applicants; confirmed victim reports4  Precision/recall by action class:5    · Auto-ban: precision primary (appeal overturn rate low)6    · Hold-for-review: recall primary for high-risk patterns7  Guardrails: time-to-first-publish for verified SMEs; legit job fill rate8  Ops: review queue SLA; auditor agreement rate910  Appeal: in-product form → evidence upload → human review <24–48h for11  paying employers; temporary distribution restore if low-risk1213  Launch gates:14    · Shadow-mode classifier for 2 weeks (no user impact) → precision check15    · Soft friction on 5% traffic → measure legit conversion impact16    · Full enforce with kill switch + on-call1718  XFN: trust ops, legal, growth (SME onboarding friction), data science,19  payments (chargeback patterns on paid posts)

1WEAK: 'Hire more moderators and ban bad users. Zero tolerance. AI will detect2everything. Success = number of bans.'34STRONG: attacker economics · layered controls · action-class precision/recall ·5false-positive cost to SMEs · appeals SLA · shadow→friction→enforce · kill switch.

1CAPSTONE MOCK PROMPT BANK (pick one)23  1) IMPROVE METRIC: Consumer fintech app — weekly funded actives flat YoY.4  2) AI FEATURE: Design AI search for a B2B knowledge base product.5  3) STRATEGY: Should a major food-delivery app launch grocery in a new country?6  4) MARKETPLACE: Gig platform — demand high, supply thin at peak hours.7  (+ optional) T&S: Prevent review fraud on a two-sided marketplace.89  TIMING (40m):10    0–3m   clarify + thesis + type-spot out loud11    3–12m  structure (tree / threat model / beachhead — by type)12    12–28m depth: options, cuts, metrics, risks13    28–35m pushbacks (cut 50%, competitor moves, cost)14    35–40m summary: decision, metrics, kill criteria1516  SELF-SCORE: rate 1–5 on each of the 10 L1 dimensions.17  REWRITE: re-record 8 minutes fixing only the two lowest scores.

1SPOKEN SENIOR ANSWER A — Local services marketplace, demand cold (~12 min)23  [0:00–1:30 THESIS + COLD SIDE]4  "Marketplace case. Supply of pros is high in three metros; demand bookings5  are weak. I will not treat both sides symmetrically. Cold side is demand —6  but get more consumers is incomplete. Demand is often weak because of trust,7  selection quality, price transparency, or time-to-book, not only awareness.8  I also check whether supply is real activity or zombie listings inflating9  vanity supply."1011  [1:30–4:00 DIAGNOSIS FRAME + DATA]12  "Frame: which side is cold; is the hot side real active pros last thirty days;13  liquidity as successful jobs over search or request; leakage off-platform;14  unit economics by side; trust via cancel and no-show; density before breadth.15  Data asks: search to contact to book to complete; time-to-first-book; pros16  active versus listed; response SLA; phone reveal and chat export signals;17  complaint reasons. Suppose: searches happen, contact OK, book rate low; quotes18  vary wildly; first-timers fear no-shows; supply subsidies wasted on idle pros.19  Mechanism: demand blocked by uncertain price and quality and weak guarantees."2021  [4:00–8:00 PLAN]22  "Non-goal: national expansion and more pro acquisition bonuses. Bet one:23  instant-book SKUs in one category and geo — fixed-price cleaning with vetted24  pros and on-platform payment to remove quote friction. Bet two: trust package —25  on-time guarantee, easy rebook, identity badge that revokes on no-show. Bet26  three: demand growth only in dense micro-geos via performance marketing and27  property-manager partnerships where SLA is feasible. Bet four: leakage28  reduction — masked comms, milestone payment release, insurance value-add.29  Defer cutting take rate as first move; it races to the bottom without fixing30  trust. Metrics: completed jobs per weekly active demander in focus geos;31  drivers search-to-book, response SLA, cancel rate; guardrails contribution32  margin and refunds; supply health via utilization. Kill if instant-book margin33  is negative after eight weeks without retention lift."3435  [8:00–12:00 RISKS + CLOSE]36  "Risk: subsidizing the hot side makes idle anger worse. Risk: instant-book37  quality variance destroys trust — mitigate with tight vetting and revoke.38  Risk: leakage remains if payment UX is worse than Venmo — mitigate with39  guarantee that off-platform cannot match. Close: cold-side demand with trust40  and price certainty; density before breadth; unit economics; utilization41  guardrail; no vanity pro bonuses."

1SPOKEN SENIOR ANSWER B — Fake job phishing on a job marketplace (~12 min)23  [0:00–2:00 THREAT MODEL]4  "Trust and safety case. Threat model: attackers post fake high-wage jobs to5  harvest resumes, government IDs, and bank details. Attacker cost is low —6  cheap accounts, stolen employer brands, copy-paste text. Defender goals: raise7  attacker cost, reduce victim harm, preserve legitimate SME hiring because false8  positives fire real jobs. Harm is financial, privacy, and platform trust."910  [2:00–6:00 LAYERED CONTROLS]11  "Defense in depth. L0 prevention: employer identity verification tiers — email12  domain, company registry, payment instrument, manual KYC for high risk; rate13  limits on new employer posting velocity. L1 detection: classifiers for phishing14  language and off-platform payment asks; graph reuse of job text, device and IP15  clusters, brand impersonation; structured seeker reports. L2 friction:16  unverified employers get limited distribution and no resume PII unlock; high-17  risk postings hold for review before apply. L3 enforcement: takedown, account18  and device bans, legal path; category kill switch if attack wave. L4 recovery:19  seeker alerts if they touched later-confirmed scams; appeal flow for false20  employer enforcement with SLA. Non-goal: one hundred percent manual review of21  every job — does not scale and destroys SLAs."2223  [6:00–9:30 FP ECONOMICS + METRICS + LAUNCH]24  "False-positive economics: blocking a real SME costs revenue and brand; missing25  a scam costs user harm and regulatory tail. Therefore high precision on hard26  enforcement bans; high recall on soft friction holds. Never one global abuse27  score for all actions. Metrics: scam-job contact rate per thousand applicants;28  confirmed victim reports; precision and recall by action class; guardrails29  time-to-first-publish for verified SMEs and legit fill rate; ops queue SLA and30  auditor agreement. Appeals: in-product evidence upload; human review within31  twenty-four to forty-eight hours for paying employers; temporary distribution32  restore if low risk. Launch: shadow-mode classifier two weeks; soft friction on33  five percent traffic measuring legit conversion; full enforce with kill switch34  and on-call."3536  [9:30–12:00 XFN + CLOSE]37  "XFN: trust ops, legal, growth on SME onboarding friction, data science,38  payments chargeback patterns on paid posts. Close: attacker economics, layered39  controls, action-class metrics, appeals, shadow then friction then enforce.40  Ban counts are not the north star — harm reduction with FP economics is."

1PRESSURE Q → SENIOR REPLY23  Q: "Why not improve both sides equally?"4  A: Symmetric plans waste capital. Subsidize or fix the cold constraint; watch5     utilization so the hot side does not churn later.67  Q: "Just lower take rate to win demand."8  A: Without trust and price certainty, lower take rate is a race to the bottom.9     Fix book friction first.1011  Q: "Zero tolerance — ban first, ask later."12  A: Hard bans need high precision. Soft friction catches more with less SME13     collateral damage. Appeals are legitimacy, not softness.1415  Q: "AI will detect all scams."16  A: Classifiers are one layer. Identity, graph, friction, enforcement, and17     appeals still required. AI without action-class metrics is theater.1819  Q: "Chicken and egg for a brand new geo?"20  A: Seed one side with ops or single-player value in a micro-geo; concierge21     first transactions; open self-serve after density and NPS gates.2223  Q: "Capstone mock — I'm out of time."24  A: Close with decision, metrics, kill criteria, and what you refused. Do not25     list ten more features.