Everybody Counts: mathematics in Cambodia

Between September 2024 and January 2025, EverybodyCounts conducted a 4-month pilot programme at PreySpeu School in Kandal Province, Cambodia, serving 229 students. In partnership with Kampuchea Action to Promote Education (KAPE) and with support from Child Fund Korea, this pilot demonstrated that hyper-localized mathematics education, culturally contextualized in Khmer language and cultural framing, significantly outperforms conventional, textbook-based approaches.

• 40% improvement in mathematics comprehension. Students using fully localized EverybodyCounts content achieved 73.1% average scores vs. 59.4% for the control group.
• 35% reduction in teacher lesson preparation time. From 6.2 to 4.0 hours per week, with automated homework marking saving 92% of grading time.
• 12% of students identified with neurodiversity indicators, enabling early intervention; accuracy only possible because students worked in native language and cultural context.

These results validate the report’s central hypothesis: that machine learning algorithms for predictive analytics and neurodiversity detection require culturally and linguistically authentic educational contexts to function accurately. When students work in non-native languages or culturally unfamiliar contexts, the resulting statistical noise prevents accurate identification of learning patterns and support needs.

Most educational technology platforms deploy content through simple translation: converting English materials into target languages without adapting the cultural context of mathematical problems. The report identifies three challenges this creates:

• Cognitive load. Students must simultaneously process mathematical concepts and navigate unfamiliar cultural references.
• Engagement deficit. Word problems referencing foreign contexts (British currency, American sports, Western food) lack relevance and fail to engage.
• Statistical noise. Machine learning algorithms struggle to distinguish genuine learning difficulties from language and cultural comprehension barriers.

The report’s example: “Sarah bought 3 apples at 50 pence each and 2 oranges at 30 pence each. How much did she spend in total?” references pence, Western names and purchasing patterns uncommon in rural Cambodia. The hyper-localized version: “Sophea bought 3 oranges at 500 riel each and 2 custard apples at 300 riel each at the local market. How much did she spend in total?”

Cambodian student names, local fruits familiar from markets and Cambodian riel currency led to more immediate recognition, reduced cognitive load and authentic engagement.

PreySpeu School sits in Kandal Province, central Cambodia, adjacent to Phnom Penh: 229 students across primary grades, a mixed rural and peri-urban socioeconomic context, and Khmer as the primary language of instruction. The report describes the school as representative of Cambodian primary schools, neither elite urban nor extremely remote rural, and large enough for statistical validity while manageable for intensive pilot support.

EverybodyCounts worked with Kampuchea Action to Promote Education (KAPE), a Cambodian NGO with established relationships with the Ministry of Education, Youth and Sport (MoEYS). KAPE reviewed and validated all Khmer translations and cultural adaptations, conducted initial professional development for PreySpeu teachers, facilitated communication with the school and local authorities, monitored implementation fidelity and worked with parents and school governing bodies. Child Fund Korea funded platform access, teacher training and evaluation, and engaged in discussions about scaling beyond the pilot phase.

Localisation followed three layers:

• Language translation. All 387 lessons translated by native Khmer speakers with mathematics education expertise, reviewed by the Cambodia National Teaching College, with terminology aligned to official MoEYS curriculum vocabulary.
• Cultural contextualisation. British pounds became Cambodian riel; pizza and burgers became rice and num banh chok noodles; Sarah and John became Sophea, Pisey and Reaksa; supermarkets became local markets and village shops; cars and buses became tuk-tuks and bicycles; Christmas and Easter became Khmer New Year and Pchum Ben.
• Pedagogical alignment. Lesson sequencing aligned with MoEYS curriculum pacing, integrated with Cambodian textbook content, with teacher notes adapted to Cambodian classroom management practices and class sizes.

All localisation underwent three-stage validation: expert review by Cambodia National Teaching College faculty, pilot testing with five PreySpeu teachers, and student focus groups of 20 students across Grades 1–6. The result: 94% of localized content required no revisions after validation; 6% underwent minor adjustments.

The study used a quasi-experimental control group design. The treatment group (115 students, Grades 1–6, six teachers) used the fully hyper-localized platform; the control group (114 matched students, six teachers) used English-to-Khmer translated materials retaining Western cultural contexts. Classes were matched within grade levels on prior-year mathematics performance (±3% average score), gender, age and teacher experience (±2 years). Implementation fidelity was tracked through weekly KAPE classroom observations (5% of all lessons), platform usage analytics, teacher self-report logs and monthly check-ins.

Assessment used adapted TIMSS Grade 4 items: 40 items selected for cultural neutrality, then localized for the Cambodian context, with parallel forms for the pre-test (September 2024) and post-test (January 2025), administered in Khmer under standardised protocols for both groups.

At baseline the groups were equivalent: treatment 52.3% (SD = 12.4%), control 51.8% (SD = 13.1%), a difference of 0.5 percentage points (p = 0.76). At endpoint: treatment 73.1% (SD = 11.2%), control 59.4% (SD = 14.3%), a difference of 13.7 percentage points (p < 0.001). Over four months the treatment group gained +20.8 percentage points against +7.6 for the control group: a difference-in-differences of 13.2 percentage points, with Cohen’s d = 0.78, a medium-to-large effect.

The analysis used difference-in-differences with multilevel modeling to account for classroom clustering, prior achievement, gender and grade. The key coefficient (treatment × post) is 13.2, p < 0.001, 95% CI [10.1, 16.3]. Robustness checks: three students withdrew (one treatment, two control) with minimal impact; results hold when excluding the top and bottom 5% of performers; effects are consistent across all grade levels.

Because the groups used different materials and delivery arrangements, the design cannot isolate cultural localisation from the platform, resources or associated support. A single-school comparison also leaves questions about replication and selection effects.

Measurement combined a hybrid time-motion study: weekly teacher time diaries across lesson planning, material preparation, assessment creation and grading; monthly KAPE observations of two teachers per group; and platform analytics for the treatment group. Baseline covered two weeks in August 2024; implementation covered the 16-week pilot.

Average weekly mathematics preparation time fell from 6.1 to 4.0 hours for the treatment group, against 6.2 to 6.0 hours for the control group: a 33% reduction on the phase comparison and 35% on the activity table. By activity: lesson planning down 52%, material preparation down 67%, assessment creation down 78%, grading down 92%. A paired t-test gives t(5) = 8.34, p = 0.0004, with Cohen’s d = 1.08; self-reported and observed time correlated at 0.81.

The report attributes the savings to pre-prepared lesson plans, ready-made materials, built-in assessments and auto-marking homework, in the teachers’ own words:

“I just need to review the plan; I don’t have to create from scratch.”

“In the past, I had to draw place value charts and number lines by hand. Now they’re all in the platform, beautifully designed.”

“This alone is worth it. Marking 35 students’ homework every night was exhausting. Now I just review the report.”

In the January 2025 survey, asked how they used the approximately two hours per week saved, all six treatment teachers reported more individual student support during class; five prepared better for other subjects; four reported professional reading and development; three reported personal rest or family time; two reported collaborative planning with colleagues. Teachers could select multiple responses.

The report’s core problem: machine learning algorithms trained to detect learning difficulties rely on identifying patterns that deviate from typical learning progressions. When students work in non-native languages or unfamiliar cultural contexts, they produce statistical noise that algorithms interpret as learning difficulties, when in reality the students are struggling with language or cultural comprehension, not mathematical concepts. Hyper-localized content in native language and cultural context eliminates this noise, allowing algorithms to detect genuine learning patterns.

Screening ran in three stages. Stage 1, automated platform analytics: response time patterns, error patterns, engagement metrics, progress velocity and task-type performance, analysed with k-means clustering, hidden Markov models, recurrent neural networks and DBSCAN. Stage 2, teacher observation: flagged students observed for four weeks (November–December 2024) against number sense, procedural consistency, spatial understanding, working memory and attention. Stage 3, specialist assessment: two educational psychologists from Phnom Penh using DSM-5-aligned diagnostic assessments adapted for the Cambodian context, blinded to platform predictions, conducted December 2024–January 2025.

Treatment group results: 21 students flagged (18.3%); 17 confirmed by teacher observation, with four not confirmed (a 19% false-positive rate between stages); 16 of 17 families consented to specialist assessment (94%); 14 students confirmed with neurodiversity indicators: dyscalculia indicators in five students (4.3%), dyslexia indicators affecting mathematics word problems in four (3.5%), ADHD indicators in three (2.6%), combined or other indicators in two (1.7%). Two assessed students fell within the normal range. The final identification rate: 14 of 115, or 12.2% of the treatment group.

Reported diagnostic accuracy against the specialist gold standard: sensitivity 89%, specificity 94%, positive predictive value 67%, negative predictive value 98%. The report interprets this as high sensitivity and very high specificity, with the 67% PPV acceptable for a screening tool where false positives are reviewed by humans before any intervention.

The same process applied to the control group flagged 34 students (29.8%, against 18.3% in the treatment group), of whom 22 were teacher-confirmed (64.7%, against 81%) and 13 specialist-confirmed (11.4%). The report’s key finding: the control group’s higher flag rate suggests cultural and language barriers created statistical noise, causing misidentification of students struggling with unfamiliar contexts. Its example: an eight-year-old Grade 3 girl flagged for slow response times and high error rates on word problems, of whom her teacher observed, “She asks what words mean (‘What is a supermarket?’ ‘What are dollars?’). Once explained, she solves correctly.” Specialist assessment identified no learning difficulty.

A preliminary July 2025 follow-up records 10 of the 14 confirmed students showing measurable improvement in mathematics performance after targeted support: individualised learning plans, parent conferences, specialist referrals where services were available, and platform accommodations such as extended time, reduced visual clutter and text-to-speech.

“Knowing specifically what each student struggles with changed everything. I can help them in ways I couldn’t before.”

The report also claims 33% more time on task, a 94% lesson completion rate and significantly more positive attitudes towards mathematics. These appear only in its conclusion, without underlying measurement detail; they remain reported claims in the data appendix.

The report credits four factors. The KAPE partnership: local NGO relationships proved essential for cultural validation, teacher training and community trust, and having Cambodian educators review localisation ensured authenticity. Intensive professional development: an initial two-day workshop plus monthly check-ins, covering Math Mastery pedagogy as well as the platform. The three-stage contextualisation process: student focus groups identified cultural references adults missed. And the multi-stage screening pathway: the human-in-the-loop approach balanced accuracy with practicality, teachers appreciated that the platform provided guidance rather than mandates, and the 94% parent consent rate suggests families trusted the process.

Four challenges, with the report’s solutions:

• Internet connectivity. Occasional outages disrupted platform access; teachers downloaded offline lesson packages and the platform synced when connectivity returned. The report recommends prioritising robust offline functionality.
• Device availability. One device per three students; rotational group work, some activities on paper, and emphasis on family devices for home learning. The report recommends a 1:2 or 1:1 device ratio for optimal implementation.
• Teacher technology confidence. Three of twelve teachers were initially intimidated by the platform; peer mentoring and a simplified teacher guide addressed it.
• Balancing research with authentic teaching. Teachers felt pressure to perform for observers and some assessment timing felt artificial; the team emphasised that the pilot tested the platform, not teachers, and introduced flexible scheduling.

Everybody Counts recommends a three-phase pathway to national implementation. Phase 1 (2025–2026): 50 schools across Kandal Province, 2,500–3,000 students, validating results across diverse school types and collecting a full academic year of longitudinal data. Phase 2 (2026–2027): 200 schools across four provinces, 10,000–12,000 students, building MoEYS capacity for national oversight and a train-the-trainer model. Phase 3 (2027–2030): all primary schools in Cambodia, approximately 7,000 schools and 1.5 million students, with MoEYS in the lead, integration with national assessment systems and sustainability through government budget.

Estimated national investment over four years: US$3.2 million for content and platform, US$4.5 million for teacher training, US$12 million for devices and connectivity (assuming 50% co-funding from schools and government), and US$2 million per year for ongoing support; a stated total of approximately US$25 million, or US$16.67 per student over four years. Proposed funding: 40% government, 50% development partners, 10% school and community contributions. Expected outcomes at national scale: 600,000+ students achieving significantly better mathematics outcomes, 15,000+ teachers with reduced workload, and 72,000+ students identified early with learning difficulties.

For a school leader or commissioner, the next questions are concrete: can the authors reconcile the workload tables, supply grade-specific assessment methods and provide the screening validation records? Replication across schools would help establish how much of the observed difference survives beyond this pilot.

The PreySpeu School pilot demonstrated, in the report’s words, that hyper-localization, combining native language with authentic cultural contextualization, dramatically improves educational outcomes and enables accurate machine learning analytics. Its key validated findings:

1. Academic impact: 40% improvement in mathematics comprehension compared to translated-only materials (13.2 percentage points gain, p < 0.001).
2. Teacher efficiency: 35% reduction in lesson preparation time (2.2 hours per week saved), with auto-marking saving 92% of grading time.
3. Accurate neurodiversity detection: 12.2% of students identified with learning support needs, only possible with native language and culture eliminating statistical noise (sensitivity 89%, specificity 94%).
4. Student engagement: 33% more time on task, 94% lesson completion rate, and significantly more positive attitudes toward mathematics.

The report calls this the Cambodia model: a partnership between an international EdTech organisation (Everybody Counts), a local educational NGO (KAPE) and a development partner (Child Fund Korea), creating a replicable framework for culturally responsive, technology-enabled learning at scale. Its closing argument: as AI-powered educational tools proliferate globally, algorithms trained on Western contexts cannot simply be translated; they must be culturally adapted, because machine learning requires authentic data to function accurately.

The tables below reproduce the source’s numeric results, comparisons and projections. Source page references use the printed page numbers, which differ from PDF page positions.