I build foundation models that hold up in the real world — and the systems that put them to work.
I turn frontier AI research into products people use. Thirteen years of building AI at Samsung Research and ARKRAY R&D: two AI products shipped to millions of devices, multiple internationally granted patents cited by Apple, Google, Meta and NVIDIA, and R&D teams of up to 13 engineers led. A PhD and 20+ peer-reviewed papers at NeurIPS, CVPR, ICLR and ECCV on robust, multimodal and agentic AI. Now building PlayPrismAI at RISE.
Now: PlayPrismAI — grounded player intelligence for game studios. Launching soon.
Meet your players before you ship.
PlayPrismAI reads thousands of real Steam reviews of games like yours and turns them into evidence-grounded player personas, predicted reactions and a ranked map of risks and opportunities for your pitch — in minutes, not months.
No team can read them all. So studios rely on intuition and learn what players feel after launch. By then it's too late.
- Finds the eight closest shipped peers on Steam by genre, mechanics and theme.
- Harvests and embeds up to 1,000 real reviews per peer, then clusters them into cited player personas.
- Each persona “plays” the pitch: predicted hook, bounce and late-game engagement.
- An authenticity gate checks synthetic reactions against real launch data — if it fails, the pipeline stops.
- Ranks risks and opportunities two ways: what design should fix, and what marketing should brag about.
Every claim cites the real reviews it came from. Ten specialist agents, a resumable LangGraph pipeline and on-premises models, wrapped in a multi-user studio platform with drill-down chat and a one-page studio memo.
Designed, architected and built by me · with support from Björn Flintberg and other RI.SE and SEAIF team members
Three questions, one agenda.
Every paper, product and patent I have worked on answers one of three questions: how to make representations reliable under shift, how to adapt foundation models to video and structure efficiently, and how to deploy AI systems people can trust.
Robustness should be learned during pretraining, not patched afterwards. Adaptive adversaries (ASTrA, ICLR 2025), camera-free perspective robustness (Möbius-MPD, ECCV 2024) and label-efficient learning for medicine and mining.
Most capability already sits inside frozen foundation models; the research problem is the small, well-placed adapter that unlocks it. DiA (NeurIPS 2026) for fine-grained video, VidTAG (CVPR 2026) for global-scale video geolocalization.
Reliability is a property of the system, not the model. Grounding, evaluation gates and the right human checkpoints decide whether AI can be deployed — from Samsung's recognition engine to PlayPrismAI's authenticity gate.
Research that left the lab.
Two AI products in market — one on millions of smart TVs, one inside a medical device — and open-source tools other people build on. Each was a team effort I led; each is backed by numbers.
Samsung ACR & Ads
Architected the visual-media recognition engine on Samsung's ACR fingerprinting — CNN and LSTM models on live TV content — plus RL-based ad recommendation. Grew from individual contributor to leading 13 engineers, with three promotions.
AUTION EYE AI-4510
Built and led the six-person team behind the on-device vision in ARKRAY's analyzer — custom ResNet classifiers and GAN super-resolution for sediment imaging — through to commercialization.
Möbius-MPD & ImageNet-PD
The ECCV 2024 method packaged for PyTorch and Albumentations, plus a benchmark for viewpoint robustness — train perspective-robust models without camera parameters.
Thirteen years between research and product — and the next decade building where they meet.
Industry work fed the questions; the PhD answered them; today's work puts the answers back into systems.
Designing and deploying foundation-model, multimodal and agentic systems on HPC infrastructure in the Sweden AI Factory — part of the EU network of supercomputing hubs. Architect and builder of PlayPrismAI.
Moved to vision-language and video models: DiA at NeurIPS 2026, VidTAG at CVPR 2026 with UCF CRCV, and an open chart-reasoning model on Qwen2.5.
Asked why models that win benchmarks fail in the world. Papers at ICLR, ECCV, ACCV and WACV; visiting researcher at UCF's Center for Research in Computer Vision with Prof. Mubarak Shah; DAAD AI-Net Fellow.
Built and led a six-person team that took on-device medical vision from 82% to 94% accuracy and into a commercial analyzer.
Content recognition and ad recommendation on millions of smart TVs; five granted patents; Innovator of the Year 2017.
Innovator of the Year
For the patent portfolio behind Samsung's AR, VR and advertising systems.
Outstanding Performance
Two company-level performance awards at Samsung Research.
DAAD AI-NET Fellowship
German Academic Exchange Service fellowship for AI research collaboration in Germany.
ML Summer School, Cambridge
Selected for the ELLIS machine learning summer school at the University of Cambridge, UK.
“Intelligence isn't memory. It's the small cognitive core that reasons beyond it. That's where I place my bet, today and for the future.”
Hosted by Prof. Mubarak Shah — and three papers since.
In 2024 I joined the Center for Research in Computer Vision at the University of Central Florida as a visiting researcher, hosted by Prof. Mubarak Shah — CRCV's founding director and one of the most influential figures in computer vision. The collaboration continues today.
The visit was featured in the Swedish Society for Automated Image Analysis (SSBA) magazine, and the Möbius-MPD work was featured by the Swedish Trans-Atlantic Researchers and Scholars Network (STARS) in 2024.
Based in Sweden, presented everywhere.
Invited talks at MBZUAI, EMBL-EBI and MIRAI; a research visit to UCF CRCV; papers presented at ICLR, ECCV, ICCV, WACV and EurIPS across five continents.
If you're building in this neighborhood — say hello.
Next, I want to build: as a research leader on frontier multimodal systems, or as a founder taking robust, grounded AI into products. If either resonates, I'd like to hear from you.
Co-founder, advisor, or trading notes on what you're seeing. Bring a concrete question; I'll have a concrete answer.
Joint projects in robustness, multimodal or agentic systems. Invited reviewer for CVPR, ICCV, ECCV, NeurIPS, ICML, ICLR, AAAI and more; ELLIS member.
On robust self-supervised learning, the gap between benchmark and deployment, and grounded agentic AI.
prakash.chandra.chhipa@gmail.com
