Case study: 24/7 content moderation for a global social platform (Bahasa + English)

The client is a global social platform that serves the Southeast Asian and broader international market. The platform's UGC volume in Indonesia is substantial: Indonesia is the fourth-largest internet market in the world with 221M+ users (APJII 2024), and a meaningful share of that volume flows through the platform's Bahasa-first lanes. The client does not authorize naming in this case study; the work is described operationally.

The platform's trust-and-safety policy covers the standard harm taxonomy: violence, sexual content, hate, self-harm, fraud, illegal goods, intellectual property, spam, misinformation, and age-related categories. The work is split by language tier (Bahasa-first, English-first, bilingual) and by severity (P0 imminent harm through P3 low-risk batch). The 24/7 follow-the-sun design gives the client continuous coverage with no more than 4-hour queue latency on the priority lane.

Bahasa-first lanes account for roughly 60% of the volume handled by the Indonesian pod: short-video reviews, comment threads, marketplace-style listings, live-stream moderation, and DM triage. English-first lanes account for the remaining 40%: global harm categories, cross-region scam patterns, fraud signals that originate outside Indonesia but surface in Indonesian-language UGC, and policy enforcement on English-language content posted by users in the SEA market.

Severity tier determines SLA. P0 (CSAM, terrorism, imminent harm) routes to a 24/7 alert queue with sub-15-minute acknowledgement and immediate escalation to a trust-and-safety lead — moderators are never the last line on a P0 decision. P1 (severe violence, hate) targets 60-minute acknowledgement during waking hours. P2 (misinformation, harassment) targets 4-hour acknowledgement. P3 (spam, low-quality) runs end-of-day batch.

Three factors drove the location decision. First, native Bahasa Indonesia fluency at scale: Indonesian moderators read regional slang, code-switching between Bahasa and English, and Indonesian cultural references that automated classifiers still miss. Mod Jakarta, Bekasi, Bandung, Surabaya, Yogyakarta, and Medan each surface slightly different slang patterns and risk contexts; the cohort draws from across these cities to cover the dialect spread.

Second, B1–B2 English proficiency (EF EPI 2025 places Indonesia at B2 Moderate, rank #80 of 123) gives the pool enough English to apply English-language harm policy consistently and write rationale notes for QA review. A C1/C2 subset of the pool handles English-only priority queues (P0 English cases, English-language fraud patterns).

Third, cultural fit: Indonesian moderators are calibrated to the SEA market in a way that offshore pools in the US or Eastern Europe are not. The same judgement calls that work for trust-and-safety on global platforms — what counts as harassment in an Indonesian comment thread, what promotional language crosses into misinformation, how to interpret a meme — are easier to make in-region.

Zipang's 5-gate funnel was applied with a gate 3 quiz specific to moderation: 20 case scenarios across P0–P3 severity tiers, scored against a published rubric covering decision accuracy, severity classification, escalation triggering, and rationale quality. The quiz includes a deliberate ambiguity case to test how the candidate handles edge reasoning.

Aggregate pass rates: gate 1 retained 24% of applicants, gate 2 retained 44%, gate 3 retained 53%, gate 4 retained 69%, gate 5 retained 76%. End-to-end, 12% of applicants reached the paid trial cohort, and 46% of trialists converted to full-time production. The 12% pass-through is higher than the France retail program's 6% because the moderation task has a wider funnel by design — more candidates are screened in, and the conversion is gated by paid trial performance rather than by interview performance.

Active moderator headcount sits at 50+ at any time, distributed across 4 shift rotations. The four rotations follow the sun: an early-morning shift in WIB (06:00–14:00) handles the European afternoon and APAC morning; a mid-day shift (10:00–18:00) handles APAC afternoon; an evening shift (14:00–22:00) handles the US East Coast morning; and a night shift (22:00–06:00) handles the US West Coast afternoon and the European morning. The four shifts overlap by 1–2 hours at each handoff, which is the operational handoff window for queue review and shift-leader briefing.

A 4-day overlap with the US/EU/APAC client teams gives the client continuous coverage: when the US East Coast team is in their morning, the Indonesian evening shift is on; when the EU team is in their afternoon, the Indonesian morning shift is on. The same work is handed off between shifts with full queue state preserved, so no case is dropped at a shift boundary.

Content moderation exposes operators to graphic and disturbing material. Treating that exposure casually is the single biggest employer mistake in this category. The program runs an explicit mental-health protocol: rotation between high-severity and low-severity queues within a shift (a 6-hour shift typically rotates every 90 minutes), capped weekly hours on graphic categories (typically 4–6 hours per day on P0/P1 queues), scheduled wellbeing breaks, and confidential counselling in Bahasa Indonesia through an external EAP (Employee Assistance Program) provider.

Opt-out paths exist for sensitive categories without career penalty. Moderators can move between severity tiers and between language lanes over their tenure; the program has explicit career paths from moderator to QA reviewer to policy specialist, all of which keep moderators in the production loop while reducing direct exposure to the most graphic categories.

Four KPIs run on the live dashboard. Precision (correct removal of policy-violating content) and recall (caught violations across the queue) are the two accuracy numbers. Time-to-decision is the queue latency by severity tier — P0 must be acknowledged in 15 minutes, P1 in 60, P2 in 4 hours, P3 end-of-day. Escalation rate is the share of decisions routed to a trust-and-safety lead for secondary review; a healthy moderation program runs escalation rates of 4–8% on P0/P1 and 1–3% on P2/P3.

The dashboard is visible to operators, shift leaders, the client's trust-and-safety program manager, and Zipang operations management. Weekly KPI reviews compare operator scores against the gold set, flag drift, and trigger retraining sessions. Monthly business reviews with the client compare Zipang's KPI sample against the client's independent QA sample; the two have tracked within 1.5 percentage points on precision and within 2 percentage points on recall across the program's life.

Two operational difficulties stood out. First, regional dialect: the same Bahasa word can carry a benign or harmful meaning depending on which Indonesian city the user is in. Slang from Medan, Jakarta, Bandung, and Surabaya each surfaced in the moderation queue, and the operator cohort was deliberately distributed across these cities so that at least one operator on every shift has native familiarity with each dialect cluster. The cohort composition is reviewed quarterly and rebalanced when one city starts to dominate the volume.

Second, evolving policy: the client updates the trust-and-safety policy every 4–6 weeks, and each update requires retraining the active production cohort. The retraining cycle is built into the program: a 60–90 minute cohort training session runs the morning after each policy update, with shadow-queue verification before operators return to live production. The discipline that holds the 90%+ sustained accuracy on a moving policy is the discipline of treating every policy update as a re-onboarding.

The 30/90-day replacement guarantee is the operational backstop when a moderator does not meet program standards. The 30-day window covers any moderator whose gold-set accuracy drops below 88% for two consecutive weeks, or who triggers an escalation event, or who fails a policy update retraining. The 90-day window covers non-performance issues that surface later: consistently slow time-to-decision, repeated QA flags, shift-coverage failures, or wellbeing-related opt-outs that don't resolve with the standard rotation protocol.

Replacement is Zipang's responsibility, not the client's. Across the life of the program, 14 moderators have been replaced under the 30-day window and 6 under the 90-day window. The current active cohort of 50+ is a mix of original hires and replacements, with the longest-tenured moderator at 38 months. The 12-month retention rate of 88%+ is calculated on the full replacement-adjusted headcount, not the original hire cohort.

The program is in its fourth contract year. Active headcount is 50+ across 4 shift rotations. Sustained decision accuracy is at 90.4% on the rolling 12-week average. The 12-month retention rate is 88%+. The mental-health program, the severity-tier rotation, and the EAP have held together through three policy-update cycles and one client-requested expansion of P0 queue coverage.

The client is not public, and the contract does not permit naming. What is disclosed is the operational shape: 50+ active moderators, 4 shift rotations, Bahasa + English, follow-the-sun 24/7 coverage, 88%+ 12-month retention, 90%+ sustained decision accuracy, 30/90-day replacement guarantee in action. The shape is reusable for any global platform with SEA Bahasa volume and English-language cross-region harm categories; the policy is the variable, the funnel is the constant.