I'm a research engineer working on large-scale multimodal AI systems that see, reason, and operate under real-world constraints. My work spans computer vision, vision-language models, reinforcement learning, distributed training, and more recently, music/audio modeling.
I enjoy building systems that move beyond benchmarks — from autonomous navigation to distributed inference pipelines — and occasionally making robots smarter (or at least slightly less confused).
- Multimodal LLMs & Vision-Language Models
- Long-context training (24K+ tokens)
- SFT & DPO
- Synthetic data generation
- Large-scale evaluation pipelines
- PyTorch DDP & FSDP
- Tensor / Pipeline / Context Parallelism
- Multi-node GPU clusters (up to 32 nodes)
- Distributed inference (100B+ models)
- 3B+ tokens/day serving pipelines
- Spark data pipelines
- Retrieval-augmented systems (RAG)
- Throughput & latency optimization
- Heterogeneous GPU deployment (AMD & NVIDIA)
-
BetterSearch
Desktop application integrating LLMs, advanced RAG, and native OS functionality for semantic file search. Supports text-to-SQL via osquery, CPU/GPU customization, and local/cloud deployment modes. -
Exo
Distributed inference framework for running LLMs across heterogeneous device clusters. Contributed AMD GPU support improvements and llama.cpp Windows backend integration. -
VidTune
Generative AI web app for tailored music creation for videos, with genre, tempo, keyword, and dynamic mixing customization. -
Unified Local-Cloud Decision-Making (UniLCD)
Residual reinforcement learning framework for cloud-edge collaboration in embodied vision systems. Presented at ECCV 2024. -
Autonomous Underwater Vehicle
Led team to Singapore AUV Challenge finals. Built navigation, perception, and embedded systems; integrated enhanced underwater vision algorithms.
-
Under Review:
WellLogBench: A Domain-Expert Curated Benchmark for Evaluating LLM Reasoning on Subsurface Well Log Data -
ECCV 2024:
Unified Local-Cloud Decision-Making via Residual Reinforcement Learning -
ETCCS 2020:
Person Re-Identification by Analyzing Dynamic Variations in Gait Sequences
-
Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) — Research Engineer I (Aug 2025 – Present, Abu Dhabi, UAE)
Training and post-training multimodal LLMs and VLMs for domain-specific reasoning (Multimodal Energy GPT).- 70B-class model SFT and DPO across multi-node GPU clusters
- Distributed inference for 235B models (3B+ tokens/day, 32 nodes)
- Synthetic data generation and large-scale evaluation pipelines
-
LossFunk (Initialize Program) — Researcher (June 2025 – Present)
Independent research on MIDAS, a modular framework for music source separation. -
Human-to-Everything (H2X) Lab, Boston University — Research Engineer / Graduate Research Assistant (Mar 2023 – Mar 2025)
Vision-language models for autonomous driving, cloud-edge routing, energy-aware navigation, reinforcement learning, and embodied AI systems. -
Robert Bosch GmbH — Research Intern (Jan 2020 – June 2020)
Open-set classification and predictive maintenance for vehicular systems. -
Center for Development of Advanced Computing (C-DAC) — Research Intern (May 2019 – Oct 2019)
Gait-based person re-identification (91.13% accuracy on CASIA-B).
-
Ottometric Inc. — Software Engineering Intern (June 2023 – Aug 2023)
Submodular optimization for dataset summarization and efficient model training. -
Ignitarium Technology Solutions — AI Engineer (Jan 2022 – July 2022)
High-performance INT8 kernel development for accelerator hardware (30% inference improvement). -
Synopsys India — Software Engineer (Jan 2021 – Jan 2022)
Apache Spark pipelines, BERT-based anomaly detection, unified production log systems. -
Thermo Fisher Scientific — Summer Intern (May 2018 – July 2018)
Embedded C++ development on ARM hardware; Linux Yocto + Qt systems.
-
Boston University
M.S. in Computer Science (Sep 2022 – May 2024)
Thesis: Efficient Vision and Language Models for Autonomous Systems -
IIITDM Kancheepuram
B.Tech. + M.Tech. in Electronics and Communication Engineering (July 2015 – June 2020)
- Website: tensorsofthewall.com
- GitHub: github.com/tensorsofthewall
- LinkedIn: linkedin.com/in/sandeshbharadwaj97
“Hurricane Sandy” wasn’t a weather event — it was me, my spiky hair, and a guitar solo on stage.
Teaching machines to see.
Scaling models responsibly.
Occasionally arguing with SLURM.