AI/ML Engineer

Role Summary The AI/ML Engineer will develop, train, fine-tune, optimise and deploy machine learning, deep learning and large language models for text, document, image and speech applications. The role requires depth in modern model architectures and training methods, rigorous evaluation, and the engineering discipline to deliver models that are efficient, reliable and secure in production.

Educational Qualification

- B.Tech./B.E., M.Tech./M.S. or M.Sc. in Computer Science, AI/ML, Data Science, Electronics or a related discipline (Must have)
- Advanced degree (M.Tech./M.S./Ph.D.) or published work in machine learning, NLP, computer vision or speech is highly desirable
- Certifications are not mandatory; demonstrated work, as listed under Minimum Evidence at Screening, shall carry greater weight
- Candidates from non-traditional backgrounds with strong verifiable work (published models, competition results or significant open-source contributions) will be considered in lieu of formal qualification

Experience

- 4+ years of hands-on experience developing and deploying AI/ML models (6+ years preferred); minimum 3 years training or fine-tuning deep learning models that were deployed in production
- Demonstrable end-to-end delivery of at least two models, from data preparation through training, evaluation, optimisation and deployment, with measured outcomes
- Demonstrable hands-on experience fine-tuning LLMs or other foundation models on domain data (supervised fine-tuning, parameter-productive fine-tuning or preference optimisation), with documented evaluation against a baseline
- Practical experience in at least two of the following: NLP and LLM applications, computer vision, document AI and OCR, speech recognition or synthesis, and predictive models on structured data
- Experience optimising and serving models on GPUs to meet production latency and throughput targets
- Experience in regulated sectors or with Indian-language data is a plus

Key Responsibilities

Model Development & Training

- Develop, train and fine-tune machine learning and deep learning models for use cases such as text classification, summarisation, entity extraction, document understanding, image matching, object detection and speech recognition
- Select architectures and pre-trained models suited to the task, the data and the deployment constraints, starting from strong baselines and preferring transfer learning to training from scratch
- Build training data pipelines covering preprocessing, feature engineering, data augmentation and synthetic data where appropriate, with versioned datasets
- Run reproducible, tracked experiments, including hyperparameter tuning, and perform systematic error analysis to improve accuracy, robustness and latency

LLM Adaptation & Applications

- Fine-tune open-weight LLMs and embedding models through supervised fine-tuning, parameter-efficient methods (LoRA, QLoRA), preference optimisation (DPO) and reinforcement learning with verifiable rewards (GRPO and its variants), where these outperform prompting
- Distil larger models into smaller,



cost-efficient models suitable for production
- Build retrieval-augmented and agentic systems with sound retrieval design, structured outputs and tool calling, integrated across commercial, cloud and self-hosted model providers through an abstraction layer
- Engineer the context supplied to models (retrieved content, memory, tool outputs and instructions), and set reasoning and token budgets for reasoning models against cost and latency targets
- Choose between prompting, retrieval, fine-tuning and smaller specialised models based on accuracy, cost, latency and data-governance requirements, and document the rationale

Computer Vision, Document & Speech AI
- Develop document-understanding pipelines combining OCR, layout analysis and vision-language models for forms, scanned documents and handwritten records
- Develop and fine-tune vision models for classification, detection, segmentation and image similarity, including embedding-based matching
- Develop and fine-tune speech recognition and synthesis models, including for Indian languages and code-mixed speech

Inference Optimisation & Serving

- Optimise models through quantisation (including FP8 and FP4 formats), pruning, distillation and compilation (ONNX, TensorRT), and apply serving optimisations such as continuous batching, prefix caching and speculative decoding
- Benchmark accuracy, latency, throughput, memory and cost per request on CPU and GPU against defined service-level targets
- Deploy models with production serving frameworks, including autoscaling, versioning and rollback, in coordination with the MLOps team
- Monitor deployed models for accuracy drift, latency, GPU utilisation and cost, and feed production failures back into training and evaluation data

Evaluation & Quality
- Define task-appropriate metrics (such as F1, AUC, BLEU, chrF, ROUGE, WER, CER and mAP), and build golden datasets, benchmarks and regression suites that run as release gates
- Evaluate LLM and retrieval systems for groundedness, hallucination and retrieval quality, using LLM-based evaluation calibrated against human labels
- Evaluate agentic systems on trajectories, tool-call correctness and task completion, and diagnose failures from execution traces
- Conduct subgroup analysis across languages, demographics, document types and acoustic conditions, and work with QA on acceptance thresholds

Responsible AI, Safety & Privacy
- Apply guardrails and test models for prompt injection, jailbreaks, leakage of personal data and unsafe outputs, in coordination with AI safety specialists
- Measure and mitigate performance disparities across cohorts, and document models through model cards,



dataset documentation and evaluation records
- Apply anonymisation and data minimisation in training and inference in compliance with applicable data protection law, and verify the licences of models and datasets used

Collaboration & Mentoring
- Work with data scientists, data engineers, MLOps and application teams to integrate models into production applications through well-defined APIs
- Critically assess emerging models and research for practical value, review the work of junior engineers, and review model deliverables from external vendors

Technical Competencies

- Programming: Python (expert level); SQL; Git; C++ or Rust for performance-critical components (desirable)
- Deep Learning Frameworks: PyTorch (primary), JAX or TensorFlow; Hugging Face Transformers, Datasets and Accelerate
- Model Architectures: transformers, mixture-of-experts and reasoning models; CNNs and vision transformers; conformer-based speech models; embedding and contrastive models
- LLM Training & Adaptation: Hugging Face TRL and PEFT, Unsloth or Axolotl; verl or OpenRLHF for reinforcement learning; supervised fine-tuning, LoRA and QLoRA, DPO, GRPO and reinforcement learning with verifiable rewards; distillation and synthetic data generation
- GPU & Distributed Training: mixed-precision training, FlashAttention, working knowledge of CUDA and GPU profiling; PyTorch FSDP or DeepSpeed for multi-GPU training (desirable)
- Inference Optimisation & Serving: quantisation (GPTQ, AWQ, FP8 or NVFP4); continuous batching, prefix caching and speculative decoding (such as EAGLE-3); vLLM, SGLang, TensorRT-LLM or NVIDIA Triton Inference Server; ONNX Runtime; llama.cpp for CPU and edge deployment
- Vision, Document & Speech: OpenCV; detection and segmentation models; vision-language models for document parsing (such as PaddleOCR-VL, Docling or the Qwen-VL family); speech recognition (IndicConformer, Whisper or newer open models) and speech synthesis
- Indic Language AI: AI4Bharat models (IndicConformer, IndicTrans2, IndicBERT, Indic Parler-TTS), Sarvam models and Bhashini services; handling of code-mixed and transliterated text
- LLM Applications & Agents: retrieval-augmented generation with hybrid search and re-ranking; vector databases (pgvector, Qdrant, Milvus or Weaviate); LangGraph or LlamaIndex; Model Context Protocol (MCP); context engineering; multi-provider integration through gateways such as LiteLLM
- Evaluation: task-specific metrics; lm-evaluation-harness, RAGAS, DeepEval, promptfoo or Inspect; agent trajectory and tool-use evaluation; human evaluation and inter-annotator agreement
- MLOps & Platforms: MLflow or Weights & Biases, DVC and model registries; Docker and Kubernetes; Ray (desirable); AWS SageMaker, Azure Machine Learning or Google Vertex AI; Langfuse or Arize Phoenix for LLM tracing
- AI Safety & Privacy: guardrail frameworks (NeMo Guardrails, Llama Guard), red-teaming tools (garak or PyRIT), anonymisation and working knowledge of differential privacy

This is a contractual job and Work from office at Delhi location

📌 AI/ML Engineer (Delhi)
🏢 National e Governance Division
📍 Delhi