- Joined
- Sep 2026
- How they joined
- Directly
- Before Cognition
- Scale AIML Research Engineer
- Status
- Current
- At Scale AI (May-Aug 2026, San Francisco) he was an ML Research Engineer in Physical AI & Robotics research, producing Augmentron, a NeurIPS 2026 project that augments real robot demonstrations.Aug 2026Augmentron: augmenting real robot demonstrations (NeurIPS 2026)
- X account @tanaytndn (created Mar 2015, ~336 followers) has the bio '🇨🇾 at penn - @cognition, @scale_ai', publicly tying the account to both Cognition and Scale AI.Nov 2026
- Co-authored arXiv:2608.13729 'Limitations of Synthetic Data Generation in Specialized Data-Scarce Domains' (Aug 2026) with Penn researchers including Edward Zhang, Marcel Hussing, Camillo Jose Taylor, Daniel A. Hashimoto and Eric Eaton, tied to his Penn research on autonomous robot triage for the DARPA Triage Challenge.Aug 13, 2026Limitations of Synthetic Data Generation in Specialized Data-Scarce Domains (co-author)
- Augmentron writes in first person: 'I realized we could represent the room with just six images: a floor, a ceiling, and four walls' - rendering generated room assets from recorded camera poses so augmented robot demos stay multi-view consistent, rather than editing each frame with an image model.2026Augmentron: augmenting real robot demonstrations (NeurIPS 2026)
- The Penn arXiv paper evaluates synthetic-data augmentation for trauma recognition inside PRONTO, Penn's robotic triage system for the DARPA Triage Challenge using aerial and ground robots. Co-authors include faculty Camillo Jose Taylor, Eric Eaton and Daniel A. Hashimoto.Aug 13, 2026
- Asked publicly 'why do VLAs suck at handling lighting condition changes?' in June 2026. Augmentron tackles that same visual-robustness problem by re-rendering rooms and table surfaces around real demonstrations.Jun 29, 2026Stupid question: why do VLAs suck at handling lighting condition changes? Why does the VLM backbone not provide robustness to a brightly vs dimly lit object?
- Research role at the University of Pennsylvania (Jan-Sep 2026): autonomous robot triage in mass-casualty scenarios for the DARPA Triage Challenge.Jan 2026
- Served about 14 months in the Cypriot military (Greek Cypriot National Guard, Jul 2022-Aug 2023) as a Staff Sergeant deployed on the demilitarized zone, specializing in anti-tank warfare, before starting Penn in Aug 2023.Jul 2022
- Retweeted Scale AI colleague @jadechoghari on automating quality control for robot-mounted and ego-camera training data at @ScaleAILabs.Jul 14, 2026
- Software Engineer intern at Maybern in New York (May-Aug 2025), working on fund-accounting software.May 2025
- Palantir Launch Fellow (Mar 2025), a residency in Palantir's Washington DC office.Mar 2025
- Penn World Scholar, Class of 2027 from Cyprus. His ISSS profile lists formal interests in Computer Science and Psychology and informal ones in Philosophy, Finance and Mathematics, and says he was 'unsure of which technological specialty I will follow'. That contrasts with the robotics and AI research path he took within two years.Feb 19, 2025
Autonomous robot triage in mass-casualty scenarios (DARPA Triage Challenge) https://arxiv.org/abs/2608.13729
Taught Data Structures & Algorithms (CIS 1210)
Physical AI & Robotics Research https://augmentron.tanaytandon.com (NeurIPS 2026)
Wrote code to improve fund accounting.
Residency in Palantir's DC office.
Deployed on the demilitarized zone. Specialized in anti-tank warfare.
Founded Cyprus' first TEDx Youth event.
one step closer to becoming a clanker https://t.co/BtngSsgtvW
Stupid question: why do VLAs suck at handling lighting condition changes? Why does the VLM backbone not provide robustness to a brightly vs dimly lit object?
im in sf for the summer trying to get robots to do useful things @scale_ai, and apparently, posting on here helps with that so when do all the cool people and opportunities start rolling in ? https://t.co/L2ga1PpJRU
Had the pleasure of hearing from @chris_j_paxton at Penn today! Biggest takeaways for me were: - “Reasoning” models are under-explored - Hard to trust evals rn. Uniform platforms (eg Unitree G1) might make this better - DexUMI seems pretty cool and useful https://t.co/IETNKqza0e
Augmentron: augmenting real robot demonstrations (NeurIPS 2026)
Limitations of Synthetic Data Generation in Specialized Data-Scarce Domains (co-author)
finally solved one of these https://t.co/1uzz4Aot30
Are your thoughts easy to compress? (from @nabeelqu) https://t.co/3qEv3uz5zR
TIL Hindu temple construction rejects modernity (corrodible steel), embraces tradition (stone), and relies heavily on devotional volunteerism. And is better off because of all of this. https://t.co/D9q6SNj1r3
COMM 3950 at Penn. Iykyk.
makes me wanna build a mario kart world model
curious what the autonomous / teleop split will be here
