Karya’s mission is to leverage the AI revolution to create dignified earning and learning opportunities for low-income communities across the world. In India, our smartphone-based digital work platform has enabled over 120,000 people to build language datasets and evaluate AI models while earning fair wages. With pilots in Kenya and Ethiopia and partnerships with leading AI labs, we are advancing language technologies for underrepresented communities. By combining dignified digital work with frontier AI innovation, Karya hopes to build a new AI economy that is equitable, inclusive, and community driven.
Karya's AI Evaluations team measures how well AI models serve people in the languages and contexts global benchmarks miss. We work with partners including Anthropic and Microsoft Research, alongside government and civil society, running evaluations on our own platform with native speakers and domain experts across India and beyond. Our benchmarks, leaderboards and reports aim to help model developers and decision-makers understand where models work well and where they fall short.
As a Research Engineer, you will design, build and run these evaluations end to end, across eval design, model deployment, statistical analysis and engineering. The work spans ASR, TTS, machine translation and chat, with newer frontiers in agentic and multimodal evaluation.
What We Do
- Benchmark design: curating representative query sets and defining rubrics that score reliably and separate models meaningfully.
- Expert input: designing elicitation and adjudication processes that turn domain expertise into consistent, scalable judgments.
- Bias evaluation: developing methods to detect gender and other social biases in LLM outputs across languages and cultures.
- Error attribution: tracing failures in multi-stage AI pipelines to the responsible component.
- Leaderboards: aggregating noisy human and automated judgments into rankings that account for annotator reliability and uncertainty.
Key Responsibilities
- Design benchmarks, rubrics and metrics, and choose between human, automated and hybrid evaluation based on reliability, cost and validity trade-offs.
- Deploy and serve models through APIs and self-hosted GPU inference, with versioned configs, logging and controls for reproducible results at scale.
- Analyse evaluation data with sound statistical methods (agreement, uncertainty, significance, disaggregation), and validate automated judges against human ground truth.
- Build and maintain modular evaluation pipelines and leaderboard infrastructure, covering sampling, inference, task packaging, scoring and reporting, with tested, documented code.
- Prototype new evaluation methods, and turn results into publications, technical reports and public releases.
Who You Are
- 4+ years in ML engineering or applied ML research, shipping production-quality systems.
- Strong software engineering, including clean, tested Python and end-to-end data and ML pipelines.
- Deep working knowledge of modern AI architectures and inference, including quality, latency and cost trade-offs.
- Rigorous grounding in probability and statistics on noisy real-world data.
- Ability to independently turn ambiguous questions into rigorous, executable technical work.
- Master's degree in Computer Science, Machine Learning, NLP or a related field preferred; equivalent industry or research experience will be considered.
Strong differentiators
- Experience building benchmarks or evaluation systems.
- Experience with human judgment data.
- Experience with speech or multilingual NLP, especially Indian languages.
- Publications