Head - AI Research

Clarus Advisers; Bengaluru; On-site; INR 4,500,000 - 7,000,000; Full time

Job summary

Clarus Advisers is seeking a senior AI research leader to design and evaluate on-device LLM prototypes for privacy-aware personal agents in edge devices. You will drive experiments, implement efficient inference pipelines, and coordinate with hardware and product teams to scale research into production-grade features.

You will lead multi-year roadmaps, publish findings, and mentor teams while advancing state-of-the-art in adaptive test-time scaling, memory consolidation, and user personalization

Qualifications

• PhD or Master's in AI field.
• 10+ years of experience in relevant research areas (LLMs, on-device/edge AI, RL, privacy).
• Strong Python expertise and C/C++ proficiency.

Responsibilities

• Design and evaluate research prototypes for on-device personal LLM agents.
• Prototype on-device inference pipelines and profile latency, memory, energy.
• Collaborate with cross-functional teams to transition research into production.
• Establish reproducible evaluation harnesses, privacy benchmarks, and profiling standards.
• Contribute to IP strategy and publications.

Skills

Python

C/C++

Shell scripting

Java/Kotlin

Swift

PyTorch

JAX

Transformers

PEFT/LoRA

DeepSpeed

vLLM

TensorRT-LLM

ONNX Runtime

llama.cpp

MNN/TFLite

Education

PhD/Master's in AI field

Tools

Docker

MLflow

W&B

Job description

Company Overview

Our client is a Research lead organisation focusing on cutting-edge software, AI, and hardware innovation.

Position Overview

• Design, implement, and evaluate novel research prototypes for on-device personal LLM agents, with emphasis on adaptive test-time scaling, privacy-preserving memory consolidation, and RL-based personalization policies.
• Conduct literature surveys, identify research gaps, and formulate hypotheses and experimental plans aligned to the PersonalLLM research agenda.
• Build and maintain research codebases for training, fine-tuning, compression, and on-device inference of LLMs; reproduce baselines and benchmark against them.
• Design and run rigorous experiments (ablations, statistical significance, privacy-leakage measurement) and document findings in internal technical reports and publications.
• Prototype on-device inference pipelines (quantization, KV-cache management, speculative/adaptive decoding, early-exit, hardware-aware scheduling) and profile latency, memory, and energy on edge hardware.
• Develop privacy-preserving user-state representations (preference embeddings, user latent vectors, skill representations) and memory-consolidation mechanisms that discard raw interactions.
• Implement reinforcement-learning controllers for personalization policies (when to retrieve, reason longer, self-reflect, or update user representations) under cost/energy/privacy budgets.
• Collaborate with cross-functional engineering, product, and hardware teams to transition research prototypes into production-grade features.

Responsibilities

• Define the multi-year research roadmap for personal LLMs, aligning it with Samsung's device-intelligence vision and the evolving PersonalLLM agenda.
• Own end-to-end delivery of research objectives from problem framing and hypothesis design through prototyping, on-device validation, and production hand-off.
• Bridge the AS-IS to TO-BE transition: move the organization beyond distillation/pruning/quantization, RAG, and LoRA fine-tuning toward dynamic compression, semantic memory compression, neural prompt compression, learned KV-cache eviction, adaptive decoding, end-to-end agentic models with self-reflection/self-verification, private continual learning, and energy-aware inference planning.
• Establish research best practices: reproducibility, evaluation harnesses, privacy benchmarks, and on-device profiling standards.
• Represent SRIB in external research communities, open-source collaborations, and academic partnerships; build a pipeline of talent through mentoring and university engagement.
• Contribute to IP strategy by identifying patentable inventions and prior art.

Skills & Experience

• PhD/ Masters in a relevant AI field (e.g., Computer Science, Machine Learning, Artificial Intelligence, Natural Language Processing, or a closely related discipline).
• Experience
• Minimum 10 years of experience in a relevant research area (LLMs, on-device/edge AI, reinforcement learning, continual learning, privacy-preserving ML, or federated learning).
• Technical Skills
• Languages & Scripting: Python (expert), C/C++ (intermediate+), shell scripting; familiarity with on-device/mobile development (Java/Kotlin or Swift) is a plus.
• Deep-Learning & LLM Frameworks: PyTorch (expert), JAX (intermediate); Hugging Face Transformers, PEFT/LoRA, DeepSpeed, vLLM, TGI, TensorRT-LLM, ONNX Runtime, ExecuTorch, MLC-LLM, llama.cpp, MNN/TFLite.
• Model Compression & Efficient Inference: Distillation, pruning, quantization (INT8/INT4, weight-only, activation-aware), LoRA/QLoRA; dynamic compression, adaptive precision, task-specific extraction. KV-cache optimization (quantization, offloading, paging, learned eviction, semantic retention); FlashAttention, speculative decoding, early exit, adaptive decoding, hardware-aware scheduling.
• Context, Prompt & Memory: RAG, summarization, sliding-window context; semantic memory compression, neural memory tokens, neural prompt compression, latent intent representation.
• Reinforcement Learning & Test-Time Scaling: RLHF/RLAIF, PPO/DPO, policy-gradient methods, reward modeling; test-time compute scaling, adaptive inference controllers, cost-aware reasoning frameworks, personalized inference schedulers.
• Personalization, Continual & Privacy-Preserving Learning: Continual/lifelong learning, catastrophic-forgetting mitigation, private continual learning, lifelong memory; differential privacy, federated learning, secure aggregation.
• Agentic Systems: ReAct, tool-calling, multi-agent orchestration; end-to-end agentic models, self-reflection, self-verification.
• MLOps, Deployment & Profiling: Experiment tracking (MLflow/W&B), containerization (Docker), CI/CD for ML, model versioning; on-device profiling of latency, memory, and energy (e.g., Android profiler, TFLite delegate tooling, NPU/GPU/DSP runtimes).
• Research Tooling: LaTeX, reproducible-evaluation harnesses, statistical analysis, visualization; strong publication and patent-writing practice.