AI Research Engineer

Eros Innovation; Chennai; On-site; INR 1,800,000 - 3,600,000; Full time

Job summary

Eros Innovation in Chennai is seeking a hands-on AI research engineer who can turn research ideas into working models. You will own experiments end to end, reading papers, reproducing results, and shipping scalable solutions across Indian data and products used by millions.

You will join and drive multiple workstreams, from image/video to language models, with a strong emphasis on reproducibility, documentation, and collaboration with research leadership.

Qualifications

• Read, reproduce and extend state-of-the-art papers in AI.
• Design and run experiments, data pipelines and ablations.
• Train and fine-tune large models on multi-GPU clusters.
• Build and curate Indic datasets with quality filtering.
• Define evaluation metrics and benchmarks against baselines.
• Write clear design docs and code for reproducibility.

Responsibilities

• Read, reproduce and extend state-of-the-art papers in generative AI, speech and LLMs.
• Design and run experiments: data pipelines, model architecture changes, training and ablations.
• Train and fine-tune large models on multi-GPU clusters using distributed training (DeepSpeed, FSDP).
• Build and curate Indic datasets, including annotation, cleaning and quality filtering.
• Define evaluation metrics and benchmarks, and compare models rigorously against public baselines.
• Optimise models for production: quantisation, distillation and serving with Triton or vLLM.
• Write clear design docs, experiment reports and code that others can build on, following our Jira and PR process.
• Contribute to papers for top conferences and workshops (CVPR, NeurIPS, ICASSP, Interspeech and similar).

Skills

Python

PyTorch

Deep learning fundamentals

Experiment design

Distributed training

Education

B.Tech/B.E., M.Tech/MS or PhD in CS, AI, EE, Maths or related field

Tools

DeepSpeed

FSDP

Megatron

TensorRT

Job description

You will turn research ideas into working models. That means reading the latest papers, reproducing them, adapting them to Indian data, training at scale, and shipping what works into products used by millions.

This is a hands-on role for someone equally comfortable in a training loop and a research paper. You will own experiments end to end and work directly with the Head of AI Research.

What You'll Work On

You will join one or more of these workstreams, based on your strengths:

• Image and video generation: Diffusion and flow-matching models for culturally grounded Indian imagery and cinematic video.
• Voice and speech: Multilingual, emotion-controllable TTS for Indian languages.
• Music generation: Multi-language music models, LoRA adaptation and serving.
• Language models: Fine-tuning and evaluating Indic LLMs against public benchmarks.
• Agentic AI: Multi-agent systems that power creative and production workflows.
• Film and media AI: Restoration, rotoscoping and masking for film archives.
• Evaluation: Benchmarks for model reliability, uncertainty and cultural accuracy.

Key Responsibilities

• Read, reproduce and extend state-of-the-art papers in generative AI, speech and LLMs.
• Design and run experiments: data pipelines, model architecture changes, training and ablations.
• Train and fine-tune large models on multi-GPU clusters using distributed training (DeepSpeed, FSDP).
• Build and curate Indic datasets, including annotation, cleaning and quality filtering.
• Define evaluation metrics and benchmarks, and compare models rigorously against public baselines.
• Optimise models for production: quantisation, distillation and serving with Triton or vLLM.
• Write clear design docs, experiment reports and code that others can build on, following our Jira and PR process.
• Contribute to papers for top conferences and workshops (CVPR, NeurIPS, ICASSP, Interspeech and similar).

Must Have

• B.Tech/B.E., M.Tech/MS or PhD in CS, AI, EE, Maths or a related field — or equivalent proven work.
• Strong Python and PyTorch; you can write a training loop and debug it without a framework hiding the details.
• Solid grasp of deep learning fundamentals: transformers, attention, optimisation, and at least one of diffusion models, speech models or LLMs.
• Hands-on experience training or fine-tuning models on GPUs, including LoRA or other parameter-efficient methods.
• Ability to read a paper and implement it from scratch.
• Good maths foundation: linear algebra, probability and calculus.
• Clear written communication you document experiments so others can reproduce them.
• At least one concrete artefact: a paper, a strong GitHub repo, a Kaggle result or a shipped ML system.

Nice to Have

• Publications at top AI, vision or speech venues.
• Experience with large-scale distributed training (DeepSpeed, Megatron, FSDP) on GPU or TPU clusters.
• Inference optimisation: TensorRT, Triton Inference Server, vLLM, CUDA kernels.
• Work with Indian languages, speech or media data.
• Experience building agents or tool-using LLM systems.
• Cloud experience on GCP (Vertex AI, GKE, GCS).
• Open-source contributions to ML libraries.

What We Offer

• Access to world-class GPU compute and a large proprietary Indian film and music catalog.
• Ownership of real research problems from the first month, not just support tasks.
• Support and time to publish at top conferences.
• A small, fast team with direct access to research leadership.
• Competitive compensation, based on experience.