Lead AI/ML Project Engineer

Bharat IT Services Limited; Dadri; On-site; INR 1,800,000 - 3,200,000; Full time

Job summary

Bharat IT Services Limited is seeking a Lead AI/ML Project Engineer to architect, fine-tune, optimize and deploy platform integration of digital systems. The role focuses on developing domain-specific SLMs for secure, low-latency environments and transitioning from cloud APIs to private, on-prem solutions.

The candidate will lead edge deployments, quantization, and evaluation with guardrails, ensuring governance and secure access across regions and devices.

Qualifications

• Must have hands-on AI/ML deployment and SLM fine-tuning experience.
• Experience with PEFT / LoRA / QLoRA.
• Experience with on-prem/offline deployment and edge computing.

Responsibilities

• Model selection and fine-tuning of SLMs for enterprise data.
• Optimization and quantization to maximize inference and reduce VRAM.
• Edge and local deployment on enterprise servers or private clouds.
• Evaluation with benchmarks and safety filters.

Skills

Power Automate

Copilot Studio

Dataverse

Custom connectors

REST APIs

Microsoft Entra

Power Platform security

Scripting basics

Tools

vLLM

llama.cpp

Ollama

TensorRT-LLM

NeMo Guardrails

Job description

Lead AI/ML Project Engineer (SLM skills and Platform Integration skills)

Job Summary:

The Lead AI/ML Engineer will architect, fine-tune, optimize, and deploy Platform Integration of current Digitalized systems in a Company” plus developing/curating Small Language Models (SLMs) tailored for high-security, low-latency enterprise environments. You will lead the transition from massive, expensive cloud APIs toward lightweight, private, domain-customized SLMs (e.g., Llama 3, Phi-3/4, Mistral) running on-premises or on dedicated edge hardware.

Key Responsibilities

• Model Selection & Fine-Tuning:Select base open-weights SLMs and apply parameter-efficient fine-tuning (PEFT, LoRA/QLoRA) and supervised instruction tuning on domain-specific client data.
• Optimization & Quantization:Compress models using 4-bit/8-bit quantization and export formats (e.g., ONNX, GGUF, Tensor RT) to maximize inference throughput while reducing VRAM usage.
• Edge & Local Deployment:Deploy quantized SLMs onto local enterprise servers, edge devices, or private clouds via runtimes like vLLM,llama.cpp, and Ollama.
• Evaluation & Guardrails:Establish benchmarks (F1, hallucination metrics, latency) and implement safety filters using tools like NeMo Guardrails.

Key Outputs desired :

Build Power Automate workflows for SLA breaches, spares delays, renewals, bank-guarantee expiries and scheduled management summaries. - Develop an authenticated Copilot Studio assistant for Microsoft Teams, with approved actions to retrieve governed business metrics. - Return answers grounded in authorised data, with reporting period, source and refresh time; handle unavailable or ambiguous data without inventing figures. - Enforce user-specific access throughout the assistant and underlying data services; test cross-region and confidential-data restrictions. - Implement approval controls for actions that change business records; maintain execution logs and exception handling. - Manage development/test/production releases, solution packaging, user documentation and training.

Must-have skills: Power Automate, Copilot Studio, Dataverse, custom connectors, REST APIs, Microsoft Entra authentication and Power Platform security/governance. Basic scripting or development ability is important when low-code tools are insufficient.

Desired Skills(May-Haves)

• Core AI/ML Stack:Advanced proficiency in Python, PyTorch, Hugging Face Transformers,peft, andbitsandbytes.
• SLM Fine-Tuning Expertise:Demonstrated hands-on experience using LoRA, QLoRA, and supervised fine-tuning (SFT) techniques.
• Inference Engines & Quantization:Direct experience with high-speed inference runtimes (vLLM, TensorRT-LLM,llama.cpp) and model compression.
• Synthetic Data & Data Pipeline Curation:Ability to generate synthetic datasets, sanitize inputs, and format instruction-response pairs.
• GPU Resource Optimization:Deep knowledge of GPU VRAM allocation, CUDA configurations, and memory reduction methods (mixed precision FP16/BF16, gradient checkpointing).