The incumbent will solve complex business problems through predictive modeling, deep learning, and generative AI research. Responsibilities span from mathematical formulation and prototype development to training large-scale models (1B+ parameters) and rigorous evaluation. This position requires deep expertise in machine learning theory, neural network architectures, and statistical modeling, with emphasis on creating novel solutions rather than infrastructure management.

CORE RESPONSIBILITIES :

- Design and architect neural network models including Transformers, CNNs, RNNs, and hybrid architectures; make decisions on layer configurations, attention mechanisms, activation functions, and connectivity patterns for optimal performance.

- Develop and implement training algorithms and optimization strategies including custom loss functions, learning rate schedules, gradient clipping, and regularization techniques to ensure stable convergence and generalization.

- Fine-tune pre-trained foundation models (LLaMA, Mistral, BERT, GPT, T5) using Parameter- Efficient Fine-Tuning(PEFT) methods including LoRA, QLoRA, Prefix Tuning, and AdaLoRA for domain-specific applications.

- Implement Reinforcement Learning from Human Feedback (RLHF) and Constitutional AI methodologies; design reward models, policy optimization algorithms (PPO, DPO), and human preference learning systems.

- Engineer high-quality training datasets through data collection strategies, cleaning pipelines, augmentation techniques,and synthetic data generation; ensure data representatives and bias mitigation.

- Design and execute comprehensive model evaluation frameworks including statistical significance testing, cross-validation strategies, benchmark dataset evaluation (MMLU, HumanEval, GLUE, SuperGLUE), and custom metrics development.

- Develop Retrieval-Augmented Generation (RAG) architectures including embedding model selection, retrieval algorithms, context integration strategies, and relevance scoring mechanisms to enhance model accuracy.

- Optimize model architectures for efficiency through knowledge distillation, model pruning, quantization-aware training,and neural architecture search (NAS) without compromising accuracy.

- Perform rigorous statistical analysis and hypothesis testing on model outputs; identify failure modes, error analysis, and edge cases requiring architectural improvements.

- Collaborate with domain experts to translate business requirements into mathematical formulations and ML problem statements; define target variables, feature spaces, and success criteria.

- Mentor junior researchers and engineers on machine learning theory, algorithmic best practices, experimental design, and research methodologies; conduct code reviews for model implementations.

- Document research findings, model architectures, training methodologies, and experimental results in technical reports; publish papers in conferences or journals and present at technical forums.

- Analyze model interpret ability and explain ability using attention visualization, SHAP values, LIME, and gradient-based attribution methods to ensure transparency in AI decision-making.

ESSENTIAL QUALIFICATIONS & EXPERIENCE :

Educational Qualifications :

- Bachelor's degree (B.E./B.Tech) in Computer Science, Engineering, Mathematics, Statistics, Physics, or related quantitative field from a recognized university.

- Master's degree (M.Tech/MS) or PhD in Machine Learning, Artificial Intelligence, Computer Science, or related field highly desirable; exceptional candidates with Bachelor's degree and significant research experience may be considered.

- Strong foundation in linear algebra, calculus, probability theory, statistics, and optimization theory essential.

Experience Requirements :

- Minimum 5-9 years of research and development experience in applied machine learning, with demonstrable expertise in designing and training neural networks.

- Extensive experience in at least two domains : Natural Language Processing (NLP), Computer Vision, Speech Recognition, Recommendation Systems, or Reinforcement Learning.

- Research Publications : First-author publications in Tier-1 ML conferences (NeurIPS, ICML, ICLR, ACL, CVPR) or journals (JMLR, TPAMI, TACL) (are preferred).

TECHNICAL COMPETENCIES REQUIRED :

Machine Learning Theory & Algorithms :

- Deep theoretical understanding of machine learning algorithms including supervised learning (SVMs, Random Forests, Gradient Boosting), unsupervised learning (clustering, dimensionality reduction, GMMs), and deep learning architectures.

- Expert-level proficiency in PyTorch (strongly preferred) or TensorFlow for implementing custom models, loss functions, and training loops from first principles.

- Advanced knowledge of transformer architectures (BERT, GPT, T5, LLaMA, Mistral) including self- attention mechanisms, positional encodings, layer normalization, and feed-forward networks.

- Mathematical optimization : Gradient descent variants (SGD, Adam, AdamW, LAMB), learning rate scheduling (cosine annealing, warm restarts), second-order methods, and convex/non-convex optimization theory.

- Statistical modeling : Bayesian methods, probabilistic graphical models, hypothesis testing, confidence intervals, and experimental design (A/B testing, factorial designs).

Deep Learning & Generative AI :

- Fine-tuning methodologies : Full fine-tuning, freezing strategies, layer-wise learning rates, discriminative fine-tuning, andParameter-Efficient Fine-Tuning (LoRA, QLoRA, Prefix Tuning, P- tuning, Adapter layers).

- Reinforcement Learning : Policy gradient methods (REINFORCE, A2C, PPO), Q-learning, actor- critic architectures, and RLHF implementation for language models.

- Generative modeling : Understanding of VAEs, GANs, diffusion models, and autoregressive generation techniques.

- Model compression : Knowledge distillation, pruning (structured/unstructured), quantization-aware training, and neural architecture search (NAS).

- Embedding techniques : Word2Vec, GloVe, FastText, contextual embeddings (ELMo, BERT), sentence embeddings (Sentence-BERT, SimCSE), and contrastive learning.

Research & Development Tools :

- Experiment tracking : MLflow, Weights & Biases, or Neptune for logging experiments, hyperparameters, and results.

- Data manipulation : Pandas, NumPy, SciPy, Scikit-Learn for data analysis and classical ML algorithms.

- NLP libraries : Hugging Face Transformers, Tokenizers, Datasets library, SpaCy, NLTK.

- Visualization : Matplotlib, Seaborn, Plotly, TensorBoard for model visualization and performance analysis.

- Version control : Git for code management; DVC for data and model versioning.

- Hardware Awareness : Understanding of GPU/TPU architecture implications for model design (memory constraints, mixed-precision training, model parallelism strategies) without responsibility for infrastructure setup.