Machine Learning Engineer, Teaching Intelligence

Overview

About Visionary
Visionary is building India's learning infrastructure layer — the AI system through which students across India learn, understand, and grow. Our AI teaches in 11 Indian languages, adapts to each student in real time, and is designed to reach 162 million K-12 students across India.
We are a 15-person founding team with early government partnerships in West Bengal and 20,000 pre-registered users. Every ML system we build is used by real students, measured on whether they actually understood — not on benchmark scores.

About the role
The ML Engineer, Teaching Intelligence owns the core AI that makes Visionary different from every other learning platform: an AI that does not just answer questions but adapts its teaching approach until a student genuinely understands.
This problem has not been solved in Indian languages at scale. You will not be implementing existing research. You will be developing the approach, building the evaluation framework to measure it, and shipping it to students within 90 days.

Responsibilities

• Design, build, and own the re-explain loop the system that generates genuinely different teaching approaches when a student does not understand
• Build the curriculum-aware understanding assessment layer going beyond right/wrong scoring to identify specific misconceptions and determine a student's readiness to progress
• Develop fine-tuning pipelines for Indian language teaching quality with a specific focus on Bengali and Hindi in the first phase
• Build evaluation frameworks for teaching quality that measure pedagogical effectiveness, not just grammatical correctness or benchmark performance
• Work with real student session data from Day 1 analyse failure modes, identify patterns, and build models that address the specific gaps you find
• Collaborate with the AI Systems Architect on model serving, latency optimization, and production deployment
• Own the full ML lifecycle for your systems: data → modelling → evaluation → deployment → monitoring
• Conduct regular testing sessions with real students your measure of success is whether students understand better, not whether model metrics improved

Minimum qualifications

• 3+ years of experience building and deploying NLP or LLM-based systems in production
• Demonstrated experience fine-tuning language models not API integration, fine-tuning — with clear understanding of what worked and why
• Experience building evaluation frameworks for model quality when standard benchmarks do not exist or do not measure what matters
• Ability to analyse model failures in production and translate observations into training or architecture improvements
• Familiarity with the challenges of low-resource or morphologically complex languages in NLP systems
• Experience moving from problem definition to working prototype independently within one to two weeks

Preferred qualifications

• Experience working with Indian language datasets or models (IndicBERT, MuRIL, Sarvam, Dhruva, or similar)
• Familiarity with RLHF, DPO, or preference-based fine-tuning methods
• Experience with RAG systems and vector retrieval for knowledge-grounded generation
• Background in or strong interest in education, pedagogy, or learning science, candidates who read education research alongside ML papers are strongly preferred
• Experience designing adaptive systems that respond differently based on inferred user state
• Prior work in EdTech, tutoring AI, or any system where the quality measure is human understanding rather than task completion