Member of Technical Staff - AI/ML Engineer

Job Overview

We're looking for an AI engineer with genuine depth across three areas that most people only have one of: classical ML, LLMs, and computer vision. Our problems don't fit neatly into one bucket. A single workflow might involve a vision-language model extracting fields from a 40-year-old scanned sale deed, a ranking model deciding which retrieved records matter, and an LLM-powered agent reasoning over the results to flag title risks.

You'll work on systems that are already in production and used by real customers making high-stakes property decisions - not research prototypes.

What You'll Work On

Document understanding at scale. Fine-tuning VLMs (Qwen, Nemotron, Kimi family models using LoRA adapters) for classification, layout analysis, and field extraction across Indian property documents - handwritten, scanned, stamped, multilingual, and frequently degraded.

Classical ML where it earns its keep. Ranking and retrieval (BM25 and learned rankers), entity resolution across noisy government records, fraud/anomaly detection, and calibration of model confidence for legal-grade outputs.

LLM agents in production. Improving our conversational agents for land-records search and title diligence - tool design, context management, evaluation harnesses, and cost/latency optimization.

Evaluation and data infrastructure. Designing annotation taxonomies, building eval sets that reflect real document distributions, and closing the loop from production failures back into training data.

What We're Looking For

6–10 years building ML systems in production, with shipped work across at least two of: classical ML, LLMs/NLP, computer vision.

Strong fundamentals - you can reason about why a model fails, not just swap in a bigger one. Comfort with the full lifecycle: data, training, evaluation, deployment, monitoring.

Hands-on experience fine-tuning open-weight models (LoRA/QLoRA, SFT, preference optimization) or training CV/document models (detection, layout, OCR pipelines).

Practical LLM engineering: prompt and context design, structured/constrained outputs, RAG, agent tool design, building evals that actually predict production quality.

Solid Python and the engineering discipline to write code teammates can build on. Experience with PyTorch and the modern inference stack (vLLM or similar) is a plus.

Pragmatism. You pick the simplest approach that solves the problem - sometimes that's a gradient-boosted tree, sometimes it's an 8B VLM with constrained decoding.

Nice to Have

Experience with Indic languages, OCR for degraded documents, or multilingual NLP.

Work on agentic systems, multi-step tool use, or LLM orchestration frameworks.

Exposure to legal, fintech, or other high-stakes domains where correctness and provenance matter.

Contributions to open-source ML tooling or published applied work.

Your First 90 Days

Days 1–30: Ground truth. Ship a small improvement to a production model or eval in week one. Read real documents and real transcripts - sale deeds, ECs, agent conversations - until you understand why this data breaks naive approaches. Own one document type's extraction quality end to end.

Days 31–60: Own a model surface. Take full ownership of one pipeline - say, an extraction adapter for a major state or the retrieval/ranking layer - including its eval set, error analysis, and a measurable quality lift you've shipped to production.

Days 61–90: Shape the roadmap. Propose and begin executing a meaningful bet - a new adapter architecture, a better eval harness, a classical-ML component that cuts cost or error - backed by evidence from your first 60 days. By now your judgment should be influencing what the team builds next, not just how.

Why This Role

Frontier applied-AI problems with no playbook - nobody has solved document intelligence for Indian land records.

Direct impact: your models decide whether a family's property purchase is safe.

Small, senior team with high ownership; you'll shape architecture, not just implement