Work

Case study 1

Natural-language search in Samsung Gallery

Find a photo by describing it.

  • Vision-language model
  • Semantic image retrieval
  • Research to product
Problem
People remember moments, not file names or tags. Keyword search cannot answer “the photo where we were all on the beach at sunset”.
My role
Team lead, architect and reporting manager. I coordinated three sites — Samsung Research in Seoul, the Mobile eXperience division in Suwon and our lab in Bangalore — to take the work from research to commercial release.
Approach
A vision-language model (VLM) for semantic image search in natural language, designed as a single unified backbone that also serves other Gallery features rather than one model per feature. We also built a machine-translation engine for Gallery search.
Outcome
Commercialised in Samsung’s Galaxy S25 series.
Recognition
  • Best Project award from Samsung Research, Seoul, for natural-language search using a VLM in Gallery on the Galaxy S25 series
  • New Valley Project of the Year (R&D), Samsung R&D Institute India – Bangalore (SRI-B) Annual Awards 2024
  • Super Tech Award (Development to Market), Q4 2024
Dr. Pinaki Bhaskar holding the Gallery Search team’s New Valley Project of the Year trophy in front of the SRI-B annual awards backdrop
Public write-up

Samsung’s own account of the feature

Case study 2

On‑device personalised knowledge graph

Personal context, built and kept on the device.

  • Knowledge graph
  • On-device ML
  • Personalisation
  • Privacy
Problem
Personalised features need to understand a user’s own data, but that data is scattered across apps — and it is private.
My role
Led the team from research through commercialisation, in collaboration with Samsung’s Big Data Center.
Approach
A user-centric knowledge graph constructed on the phone from the phone’s own data. Published building blocks: on-device attribute extraction from unstructured text (ICON 2021), multi-hop relation-aware representations for inductive knowledge graphs (ICON 2023), and POP-VQA (WACV 2024) — an on‑device visual question answering system that personalises its answers with the user’s knowledge graph, improving accuracy by 36% absolute on KVQA 1‑hop questions.
Outcome
Commercialised starting with the Galaxy S21 series. A granted US patent (US11989193B2) covers rewriting search queries using the user’s context and personalised knowledge graph.
Recognition
Samsung-internal global Best Project award, Mobile Developer Conference 2021 (Mobile R&D Office, Korea).
Links

Case study 3

Multimodal LLMs and Large Action Models for mobile agents

From understanding the screen to acting on it.

  • Multimodal LLM
  • Large Action Models
  • Agentic AI
  • Current research
My role
I lead the development of multimodal LLMs and Large Action Models to power agentic AI capabilities in mobile applications, including an on‑device multimodal LLM.
Related
Related generative-imaging work: a pending US patent application (US20260099969A1) on adding a missing person to an already-captured photo with a suitable pose and placement.
Links

Earlier at Samsung

  • Multimodal Story / Memory creation in Gallery
  • Natural expression generation for a social robot
  • On-device Smart Reply for Galaxy Watch
  • Smart assistant for mobile devices

Leadership

Until 2025 I managed the Advanced Technology Lab in Samsung’s Advance Research Group while staying hands-on as its architect and technical lead; since 2026 I have been one of the lead GenAI research scientists and architects in the Visual GenAI Group of the Visual Intelligence Team. I still design the systems my teams build.

Lab & people
At the lab I was the reporting manager for multiple teams and projects. I mentor engineers and researchers in adopting new ML methods, take part in recruitment, and serve on the People Task Force.
Direction
I set the lab’s technical direction from a close read of where the technology, our competitors and the market were heading, and I contribute to strategy for the wider research organisation. I also review global project proposals.
Cross-site delivery
Programmes run across Bangalore, Seoul and Suwon, with research, product and data organisations at the same table.
Growing the community
Jury member at Samsung Solve for Tomorrow 2024, Samsung India’s national innovation competition for students, and an organiser and mentor for the programme; organiser of the sponsored M.Tech. programme, hackathons and trainings. Member of the Board of Studies, Dept. of Computer Science & Engineering, National Institute of Technology Meghalaya (2024 – present).