Case study: 432 Indonesian annotators for a 100+ hypermarket retail network in France

The client is a French retail group operating 100+ hypermarkets across mainland France and overseas territories. Their internal computer-vision team had built a shelf-edge CV model in-house; the bottleneck was not the model — it was the labeled training and evaluation data feeding it. With 300,000+ active SKUs across private label and branded items, and price tags refreshed daily across thousands of stores, the dataset needed continuous high-volume, high-accuracy human review.

The client had tried internal labeling teams and two prior external vendors. Internal teams could not scale to the daily volume without headcount expansion that exceeded budget. The first external vendor delivered acceptable accuracy on simple cases but missed the price-tag and promotion edge cases that mattered most. The second vendor was cheaper but accuracy slipped below 85% during ramp, triggering the client to re-bid the program.

Zipang was the third vendor brought in, in 2022, and the program has run continuously since. This case study documents the operational shape, the methodology, and the honest results — including what did not work.

Retail-shelf CV annotation is unusually hard for three reasons. First, the SKU universe is enormous — the client's catalog runs to 300,000+ unique product-pack combinations, and the model must distinguish between, for example, two sizes of the same product with different promotional stickers. Second, price tags change every day: a hypermarket price team may update 20,000+ price tags across the network on a single Tuesday, and the labeling queue must absorb that daily churn. Third, OCR on price tags in hypermarket conditions is genuinely ambiguous: glare, partial occlusion, curved tags, and multiple overlapping promotions force a human to make judgement calls the model cannot.

The hardest subset, by error rate, was promotional and bundle stickers — temporary price reductions, multi-buy offers, and loyalty-card-specific prices. The first 60 days of the program spent a disproportionate share of QA reviewer time on exactly these cases. By month four, the team had built a gold set of 2,000+ promotion-edge-case labels that became the spine of the weekly calibration cycle.

Three factors drove the location decision. First, English proficiency at B2 CEFR is the dominant operational language for SOPs and client communication; Indonesia's EF EPI 2025 placement at B2 (Moderate) puts the production pool inside the B2 band, and Zipang's 5-gate funnel filters to candidates who clear an internal English reading and SOP-comprehension test. Second, French familiarity at working level is present in a subset of the Indonesian talent pool — university French programs, francophone media exposure, and a small but durable community of Indonesian graduates from French-speaking institutions. Third, time-zone overlap with Paris (UTC+1 / UTC+2 in summer) is partial but workable: an Indonesia morning shift (07:00–14:00 WIB) lands inside the European afternoon, and a follow-the-sun handoff to a Jakarta evening shift gives the client 12+ hours of effective overlap per day.

The scale factor is the fourth reason. The 432-operator ramp required absorbing 4–5× the typical Indonesian annotation cohort within 6 months, and Indonesia's 280M+ population (BPS 2024) made that ramp feasible without saturating any single metro labor market. The cohort was distributed across Jakarta, Bandung, Yogyakarta, Surabaya, and Medan.

Zipang's 5-gate funnel — CV relevance scan, async English and SOP-comprehension screening, role-specific quiz, structured video interview, paid trial task — was applied with a single modification: gate 3 (the role-specific quiz) was a French retail-labeling quiz built from 200+ representative shelf images and price-tag cases, scored against a rubric that mirrored the client's own gold set.

Aggregate pass rates: gate 1 retained 22% of applicants, gate 2 retained 41% of gate 1 survivors, gate 3 retained 58% of gate 2 survivors, gate 4 retained 67% of gate 3 survivors, gate 5 retained 78% of gate 4 survivors. End-to-end, roughly 6% of applicants reached the paid trial cohort; 48% of those trialists converted to paid production. The same funnel that produces Zipang's other quality programs (Transperfect–Dataforce at 90%+ sustained accuracy) was reused here — the rubric changed, the funnel did not.

Of the 432 professionals onboarded across the ramp, 208 sit in active production at any given moment. The remainder are in shadow queues, training cohorts, paid trial status, or rotating off-shift for the 4-day follow-the-sun overlap with the client's Paris team.

The 208 / 432 ratio is not attrition — it is the design. Retail-shelf CV annotation is a high-volume, high-judgement task, and the production cohort is intentionally kept below 50% of total onboarded to absorb training rotations, illness, and the deliberate 30% off-production time used for QA feedback loops, gold-set calibration, and cross-training on new SKU categories.

Production throughput runs at 3.4M tasks per month at steady state, which works out to roughly 16,000 tasks per production-operator per month, or about 600–700 per shift. Each task is a single image or short clip with a defined labeling rubric and a decision time budget of 8–45 seconds. Microsecond-level KPI tracking means every label, every reviewer intervention, and every re-annotation is timestamped in a way that lets the operations team slice the data by operator, by SKU category, by store, and by edge-case type.

The microsecond granularity is operational, not marketing. It is the difference between knowing 'a reviewer disagreed on 200 labels yesterday' and knowing 'reviewer 7 disagreed on 200 labels between 14:00 and 16:00 WIB, on the dairy-bagel SKU cluster, after a gold-set update at 13:45.' That level of attribution is what makes the weekly calibration cycle work.

Sustained production accuracy has been above 90% since month 4 of the program. The discipline that holds it is a weekly gold-set calibration: a frozen set of 500 representative cases is re-labeled by the entire active production cohort every Friday, scores are computed against the gold answer, and individual operators whose accuracy drifts below 88% on the gold set are pulled into a 1:1 retraining session before they touch live production again the following week.

Client-side QA samples 5–8% of production labels every week on an independent rubric. Client and Zipang scores have tracked within 1.5 percentage points on the weekly sample since the program stabilized, which is the basis for the 'sustained' qualifier on the 90% number — it is not a one-week spike, it is the rolling 12-week average.

Months 1–2: SOP design and gate 3 quiz construction with the client's Paris team. Gate 1 and gate 2 ran on a standard funnel; gate 3 was new for this program. 60 candidates were advanced to the first trial cohort. By the end of month 2, 28 of the 60 had converted to production.

Months 3–4: Ramp from 28 production operators to 120. Weekly gold-set calibration began. The first promotion-edge-case gold set was assembled (2,000+ labels) and became the spine of QA feedback.

Months 5–6: Ramp from 120 to 208 in production. Throughput stabilized at 3.4M tasks per month. Sustained 90%+ accuracy was first recorded in week 14. Since month 6, the program has been in steady state with 432 total deployed (training + production + rotation) and 208 in active production at any time.

The client measures three things. First, precision and recall on its own held-out test set, refreshed quarterly. Second, throughput: tasks delivered per day, per week, per month, against the agreed SLA. Third, queue SLA: time from image arrival to labeled output, with a contractual 24-hour ceiling on the standard lane and a 4-hour ceiling on the priority lane for new SKU introductions and price-tag refreshes.

Zipang's internal KPIs (per-operator accuracy, throughput, rework rate) are shared weekly with the client's Paris program manager, and the client's own QA sample is shared back with Zipang monthly. The two-way visibility is the trust mechanism that lets the program scale without re-bidding.

The first design of gate 4 (the video interview) required candidates to demonstrate conversational French. Pass-through dropped to 23% — too low to feed production ramp at the rate the client needed. French at conversational level is a smaller subset of the Indonesian talent pool than French at reading-and-SOP-comprehension level.

In month 3, the program pivoted: gate 4 dropped the conversational French requirement and gate 3 added an English-reading test of French-language SOPs (labels, promotional copy, edge-case documentation). For live French-language cases — promotion copy that needed a native speaker to confirm tone — a small specialist sub-pool was built, but the bulk of the labeling work was done in English with the client's French-language SOPs translated and reviewed by AI translation plus a bilingual reviewer.

The pivot improved ramp speed without measurable accuracy loss. By month 5, the English-plus-translation design was the default, and the conversational-French sub-pool was reduced to a 12-operator specialist team for tone-sensitive cases only.

The program is in its third contract year. The active production cohort of 208 has been stable, and the 432 total-deployed figure has been the steady-state number for 18 months. Two new product categories have been added since launch (a private-label line and a fresh-produce line), each on a 60-day ramp using the same gate structure. Sustained production accuracy is at 91.2% on the rolling 12-week average, and the client has not re-bid the program since the original contract.

The program is documented here in operational form — the client is not public, and Zipang does not disclose client identities without permission. What is disclosed is the operational shape: 432 deployed, 208 in production, 3.4M tasks per month, 90%+ sustained accuracy, microsecond-level KPI tracking, weekly gold-set calibration, 6-month ramp. The same shape is reusable for other retail-AI and shelf-edge CV programs; the rubric is the variable, the funnel is the constant.