Spanish-language content moderation and classifier training for a Latin American platform

The client operates a social and user-generated content platform with a substantial user base across Mexico, Central America and the wider Latin American region. Content volume is high and predominantly Spanish, with strong regional variation in slang, humour and cultural reference.

The platform's automated moderation performed poorly on Latin American Spanish. The toxicity classifier, trained largely on English and Castilian data, achieved 73.0% precision and 62.0% recall on Latin American content. Low recall meant harmful content persisted; low precision meant benign content was removed, generating complaints and accusations of arbitrary censorship.

The failure was concentrated in exactly the content that matters most. Regional slang, coded language and context-dependent insults were systematically missed, while vocabulary benign in one country and offensive in another was handled inconsistently.

Response time was compounding harm, with median time from report to action at 43 minutes and exceeding four hours during spikes.

Moderator wellbeing had become a retention crisis, with annualised attrition above 70% under the previous arrangement, producing a permanently inexperienced workforce and degraded decision quality.

Corpshore Mexico proposed Mérida for its Latin American demographic reach, its stable and low-attrition workforce and its quality of life, which for demanding moderation work is a genuine wellbeing advantage.

The wellbeing framework was decisive. Corpshore's documented approach, mandatory rotation off high-intensity queues, scheduled decompression, on-site counselling and peer support, was assessed as materially more developed than competing bids, and Mérida's naturally low attrition reinforced it. The client had concluded its attrition problem was the root cause of its quality problem.

Sixty-four moderators in Mérida drawn from across Latin American nationalities, eight AI training specialists on classifier improvement, five trust and safety leads and two wellbeing counsellors on site. Coverage is 24 hours on rotating shifts. All moderators complete a six-week onboarding covering policy, regional context, decision frameworks, escalation and wellbeing practice before handling live content.

Regional decision routing. Content routes to a moderator familiar with the originating region's language and cultural context rather than a generic Spanish queue.

Moderator decisions as training data. Every decision, with its regional context and reasoning code, feeds the classifier training set. Six retraining cycles were completed.

AI-augmented triage, not replacement. The classifier pre-scores and prioritises rather than acting autonomously above a narrow high-confidence band, keeping human judgement on ambiguous content.

Wellbeing as an operating requirement. Rotation limits, decompression, on-site counselling and peer support are scheduled and enforced. Mérida's low baseline attrition compounds the benefit.

Classifier precision rose from 73.0% to 94.2% and recall from 62.0% to 93.3% across six retraining cycles.

Median time from report to action fell from 43 minutes to 4 minutes, sustained through spikes because AI triage absorbs routine volume.

Moderator attrition fell from over 70% to 18%, the lowest the client had achieved anywhere, and decision accuracy rose from 80% to 96%.