Research Core AI Scientist – Foundation Models / Deep Learning

Research Scientist – Core AI - Lead

Locations: Bangalore / Delhi-NCR / Trivandrum / Mumbai (Onsite)

Experience : 10+ years

What you’ll do

• Invent new methods in deep learning/foundation models (architectures, training objectives, optimization, alignment, evaluation).
• Own the research cycle end-to-end: literature review → hypothesis → method design → experiments & ablations → paper → open-source artifacts → transfer to product teams.
• Build rigorous evaluation : benchmarks, baselines, statistical tests, reproducibility (seeds, configs, logs).
• Collaborate with engineers to scale training & inference (distributed training, profiling, efficiency).

Must-have (strict)

• PhD (preferred) or MS (Research) in AI/ML/CS

• First-author, peer-reviewed papers proposing new models/algorithms/objectives in top venues (NeurIPS/ICML/ICLR/ACL/EMNLP/TMLR/JMLR) within the last 3–4 years.

• Demonstrated from-scratch development for tabular/structured problems: e.g., novel sequence/tabular transformers, time-series objectives, structure-aware pretraining, new loss/regularizers, or agentic reasoning/evaluation methods beyond RAG/prompting.

• Strong math & systems: linear algebra, probability, optimization; PyTorch or JAX, distributed training, experiment tracking; clean research code and ablation hygiene.

Agentic AI (required exposure)

• Research experience with agentic systems (single or multi-agent) covering planning, tool-calling, self-critique/evaluators, safety/guardrails , or learning-to-reason —with novelty beyond wiring APIs.
• Evidence of method contributions (e.g., new planner/objective, coordination mechanism, evaluator metric) on structured/tabular tasks.

Nice to have

• Work on efficiency (flash-attn, quantization, LoRA/QLoRA, distillation), alignment (RLHF/DPO/ORPO), or multimodal (vision-language/audio).
• Experience designing scaling laws , data curriculum, or safety/robustness evaluations.

We are not looking for

• Computer-vision-heavy profiles (image/video-centric portfolios).
• Roles limited to prompting/RAG/agent wiring without novel methods or first-author papers.
• Only application/survey papers; no code or ablations to reproduce claims.

Interested candidates can share resume to sandhia@hiresquad.in