We are hiring a Research Engineer to train and improve reasoning models through supervised fine-tuning (SFT) and reinforcement learning (RL).

You will own the workflow from dataset preparation and baseline evaluation through SFT, reward design, RL training, and evaluation. You should have practical experience using training frameworks and modules, modifying training code, debugging experiments, and delivering measurable improvements in reasoning performance.

Responsibilities

- Build end-to-end pipelines to train reasoning models on tasks such as mathematics, coding, logical reasoning, and tool use.
- Prepare and curate demonstration datasets, reasoning examples, preference data, and tasks with verifiable answers.
- Implement SFT using appropriate tokenization, chat templates, loss masking, sequence packing, and training configurations.
- Run full fine-tuning and parameter-efficient training using LoRA or QLoRA.
- Implement and tune RL methods such as GRPO or PPO, including rollout generation, reward computation, and policy updates.
- Develop reward functions and verifiers for correctness, execution success, and task completion; identify reward hacking and unreliable reward signals.
- Integrate efficient generation engines into the RL training loop.
- Evaluate reasoning quality using held-out tasks, benchmarks, ablation studies, and failure analysis.
- Debug training instability, numerical errors, memory constraints, and throughput bottlenecks.
- Maintain reproducible experiments with versioned datasets, configurations, checkpoints, and training metrics.
- Scale successful experiments from single-GPU runs to distributed training.

Required qualifications

- Minimum 2 to 5 years of experience in AI/ML engineering, 2+ years of hands-on experience developing and training deep learning models,



including practical experience with reasoning-model post-training.
- Direct experience with both SFT and online RL, with personal ownership of training implementation, execution, debugging, and evaluation.
- Strong Python and PyTorch skills, including the ability to modify training loops, loss functions, data pipelines, and model configurations.
- Practical experience with Transformers, Datasets, TRL, and PEFT, or equivalent training frameworks and modules.
- Experience using SFTTrainer and GRPOTrainer, or equivalent components, to implement supervised and reinforcement learning workflows.
- Experience with at least one online RL algorithm, such as GRPO or PPO, and understanding of policy gradients, advantage estimation, reward normalization, and KL regularization where applicable.
- Experience developing reward functions or verifiers for reasoning tasks, such as answer checking or code execution.
- Experience with LoRA/QLoRA or full fine-tuning, including understanding of their memory, compute, and model-quality trade-offs.
- Experience training on NVIDIA GPUs using mixed precision, gradient accumulation, gradient checkpointing, and checkpoint recovery.
- Ability to design controlled experiments and demonstrate improvements over an SFT baseline on held-out reasoning tasks.

Preferred qualifications

- Experience with verl, OpenRLHF, or NeMo RL for reasoning-model reinforcement learning.




- Experience with vLLM or SGLang for productive rollout generation.
- Experience with Accelerate, DeepSpeed, FSDP, or Megatron for distributed training.
- Experience with reinforcement learning with verifiable rewards (RLVR), reward modeling, and RLHF.
- Familiarity with preference optimization methods such as DPO, alongside online RL experience.
- Experience with synthetic reasoning data, rejection sampling, distillation, and curriculum design.
- Experience training models for multi-step reasoning, multi-turn tool use, or agent tasks.
- Familiarity with evaluation tools such as lm-evaluation-harness, task-specific verifiers, and coding evaluation suites.
- Experience with Weights & Biases or MLflow, Docker, and cloud GPU environments.
- Relevant open-source contributions, research publications, or demonstrable model-training projects.

Expected outcomes in the first 90 days

- Establish a reproducible SFT-to-RL pipeline with baseline evaluations.
- Train an SFT checkpoint that improves performance on agreed reasoning tasks.
- Demonstrate additional gains through RL, supported by held-out evaluation and regression analysis.
- Document the training recipe, reward design, compute requirements, results, and next experiments.

Application requirements Please describe one SFT project and one RL project you personally implemented, including:

- Frameworks and training modules used.
- Model size, dataset scale, and GPU setup.
- Your changes to the training pipeline.
- RL algorithm, reward design, and rollout approach.
- Baseline and measured reasoning improvements.
- A training failure or performance bottleneck you diagnosed and resolved.

Anonymized project descriptions and approximate ranges are welcome.

📌 Senior AI/ML Engineer (Delhi)
🏢 Insaito Software
📍 Delhi