Work we're proud of.
Three production deployments. One coming soon. Every metric below is measured by the client, not by us.
Hindi RLHF at Scale
Preference ranking across 500 challenging Hindi prompts, spanning urban, rural and code-mixed inputs — for a foundation-model lab's alignment release.
Problem
Off-the-shelf RLHF data was failing Hindi users.
The client had scaled English-origin preference data well, but Hindi response ratings from generalist annotation vendors showed high disagreement (Kappa < 0.55) and missed cultural context — responses about rural banking, Ayurveda and local transit were being mis-ranked.
Approach
Native-paired reviewers + cultural-context rubric.
We assembled 24 Hindi-native reviewers (14 urban, 10 rural/small-town), built a 12-point cultural-context rubric, and ran every single task through four QA layers before hand-off.
4-Layer QA Breakdown
Results
TrainPlex's Hindi preference data was the first external dataset our eval pipeline graded above our internal team's work. It shipped.
Multi-Language Benchmark
Parallel evaluation across Hindi, Tamil, Bengali, Telugu and Marathi — per-language error typology for a global AI platform's Indic launch.
Problem
A single quality number hid per-language failure modes.
The client's Indic eval gave one aggregate score across five languages — useful for marketing, useless for engineering. Tamil performance was masking a 9-point drop vs Hindi; Bengali nominal forms were silently failing.
Approach
Per-language teams. Per-language rubrics.
We staffed five independent native-speaker pods, ran 100 parallel tasks per language, and delivered a 22-dimension error typology report — not just a score, but a diagnosis.
Results per language
The per-language error typology changed how we thought about our Indic launch. We went from one dashboard to five, and our user satisfaction scores followed.
Hindi Prompt Engineering
200 adversarial Hindi prompts across 5 domains — designed to expose LLM blindspots an English-origin eval would miss entirely.
Problem
Models passed English safety evals and failed in Hindi.
The client had deployed a model that was SOTA on English harm benchmarks but generating unsafe Hindi output in production. Standard jailbreaks didn't translate — but Hindi-specific adversarial patterns did.
Approach
Domain × linguistics matrix. Authored by experts.
We commissioned prompts across legal, medical, finance, education and civic domains — each authored by a credentialed domain expert and adversarially reviewed by a linguistics specialist for script, register and code-mixing attacks.
Results
We thought we were ready to ship in Hindi. TrainPlex's prompts showed us we weren't — and gave us the exact failure modes to fix before launch.
Next case study loading…
Multilingual audio annotation at scale. Publishing Q2 2026.