AIMD: AI Model Developer (SLM Specialist)

IAIRO; India; In loco; INR 2.000.000 - 3.200.000; Tempo pieno

Descrizione del lavoro

Indian AI Research Organisation (IAIRO) is seeking a hands-on AI Model Developer who builds models from the ground up, not just API calls. You will work across the full lifecycle from data curation to specialized fine-tuning, aiming to create compact, multimodal, neuro-symbolic models as part of a broader agentic framework.

You will collaborate with world-class AI engineers and infrastructure experts, tackling architectural design, distributed pretraining, and efficient inference.

Competenze

• Experience building at least one language model from scratch.
• Strong knowledge of training stability and hyperparameters.
• Experience with data curation and pretraining pipelines.

Mansioni

• Architect efficient SLM architectures optimized for latency and memory.
• Lead end-to-end pretraining, including data deduplication and tokenization.
• Perform advanced fine-tuning with RLHF, DPO, or PPO.
• Apply quantization and pruning for edge deployment.
• Develop rigorous benchmarks for domain-specific tasks.

Conoscenze

Model Building

Framework Proficiency

Pretraining Knowledge

Efficient Fine-Tuning

Scaling Laws

Strumenti

PyTorch

JAX

Hugging Face

Transformers

Accelerate

PEFT

TRL

Descrizione del lavoro

We are looking for a hands-on AI Model Developer who doesn't just call APIs, but understands the architecture under the hood. The ideal candidate has "dirty hands" from building Small Language Models (SLMs) from the ground up- someone who understands that efficiency often beats sheer parameter count.

You may either take ownership of or play a critical team role across the full model development lifecycle, from raw data curation and custom pretraining to sophisticated fine-tuning for specialized downstream tasks. While you need not be an expert in every aspect of model development and deployment, the innovative effort at IAIRO involves building a new class of AI models that are compact (i.e., small or right-sized), custom (i.e., task- or domain-specific), multimodal (i.e., capable of processing diverse sensor and visual data), neuro-symbolic (i.e., not purely data-driven but also incorporating structured domain knowledge and, where appropriate, domain expert feedback through techniques such as RL), and part of a broader composite AI system (i.e., an agentic framework that integrates multiple models, internal systems, and external tools) to execute complex tasks.

While possessing most of the following capabilities would be impressive, you will have access to some of the world’s top AI engineers with experience working at leading AI companies, as well as experts who can support the effective use of infrastructure (e.g., from Nvidia).

Key Responsibilities

• Architectural Design: Design and implement efficient SLM architectures (e.g., Transformer-based, MoE, or State Space Models) optimized for specific latency and memory constraints.
• End-to-End Pretraining: Manage the pretraining pipeline, including data deduplication, tokenization strategy, and the orchestration of compute clusters for distributed training.
• Advanced Fine-Tuning: Execute SFT (Supervised Fine-Tuning) and alignment techniques (RLHF, DPO, or PPO) to steer model behavior.
• Optimization: Implement quantization techniques (e.g., bitsandbytes, AWQ, GGUF) and pruning methods to deploy models on edge devices or other constrained environments.
• Evaluation: Develop rigorous benchmarking suites beyond standard benchmarks to validate model performance on domain-specific tasks.

Required Skills & Experience

• Model Building: Proven experience building at least one language model from scratch (not merely fine-tuning a Llama 3 or Mistral checkpoint).
• Framework Proficiency: Deep expertise in PyTorch or JAX, along with the Hugging Face ecosystem (Transformers, Accelerate, PEFT, TRL).
• Pretraining Knowledge: Solid understanding of training stability, weight initialization strategies, and hyperparameters such as learning rate warmup and weight decay.
• Efficient Fine-Tuning: Mastery of Parameter-Efficient Fine-Tuning (PEFT) methods, particularly LoRA and QLoRA.
• Scaling Laws: A grounded understanding of the relationship between compute, dataset size, and parameter count.

Bonus Qualifications

• Hardware Awareness: Experience optimizing kernels with Triton or CUDA to extract additional performance from GPUs.
• Data Engineering: Experience building high-quality synthetic data pipelines to improve model reasoning.
• Deployment: Familiarity with inference engines such as vLLM, TGI, or TensorRT-LLM.

Why This Role?

You won’t just be a user of AI; you will be an architect. This role is ideal for the engineer who finds greater satisfaction in optimizing a 1B-3B parameter model to punch above its weight class than in simply prompting a massive closed-source LLM.

You will be part of a highly innovative research team within an organization focused on sovereign AI, employing top researchers, engineers, innovators, and entrepreneurs-in-residence. Before applying, please thoroughly review: