Jalaj Upadhyay
@jalajupadhyay
Asst. Professor @Rutgers. Previously, privacy researcher @Apple, postdoc at @JohnsHopkins, @PennState, and a PhD student @UWaterloo
I am excited to share our new monograph that dives deep into correlated noise mechanisms, one that enables real-world, production-grade differential privacy at scale, without compromising learning performance. The project started with a discussion over lunch in Google's Canopy…
Just learned that Arsene Wenger could not sign Ronaldinho in 2001 for Arsenal because of the UK VISA rules for non-EU citizens. 4 years later, he led Barca to defeat Arsenal in the UCL finals.
Beyond my @icmlconf presentations, I'm also open to discussing career opportunities. As I'm completing my PhD this year, I'm actively exploring postdoc and research scientist positions. If you're attending ICML and are hiring or know of relevant openings, I'd love to chat!
I'm attending @icmlconf next week and will be presenting two posters—both focused on asynchronous distributed learning icml.cc/virtual/2025/p… I'd love to discuss the work with anyone interested, so feel free to stop by and chat.
U.S.R. Murthy and now Ray Laflamme! @UWaterloo lost two great pioneers of their fields in one summer :(
I am not sure why @overleaf needs to track my IP address. For a privacy researcher, this seems like an over-reach.
Even the deer use pedestrain crossing to cross the road at @iitmadras :)
A 22-year-old died because he could not afford asthma medication! One can get that for 500 INR (almost 90 times less) in India, not 500+ USD as in the US. nbcnews.com/health/health-…
Two nice results by my colleagues within the last week: An Inversion Theorem for BLT (arxiv.org/abs/2504.21413) and on the Space Complexity of Online Convolution (arxiv.org/abs/2505.00181).
Stunning, high-res images. Zero cloud. BlackBird by Decompute runs entirely on-device—yes, on your MacBook (16 gb RAM) Offline. Private. Fast. AI image generation just got personal: decompute.run/macosform #OnDeviceAI #BlackBird #Decompute #EdgeAI
Breathe New Energy Into AI @DecomputeAI is reimagining the cost of intelligence. LaserTune and Aurora cut energy use by up to 80%, saving over 600,000 kWh/month—all while running on your device. No datacenter. No compromise. Just pure performance. Sustainable. Local. Private.…
I wanted to share some metrics about Decompute and our mission with sustainability. Decompute’s sustainability edge lies in its radically minimized resource footprint for fine-tuning and inference, which when scaled, translates into massive savings across energy, carbon…
🚀Over 2x faster inference and significantly lower memory usage with Aurora — Decompute’s optimized Softmax replacement. It is also backward compatible with most pre-trained models. Space and compute time aren't just metrics on resource-constrained devices — they're…
On-device. Private. Fast - With @DecomputeAI I will be representing @DecomputeAI at the AI User Conference on April 17 in SF, alongside wonderful fellow panelists: - David Kanter, Founder of MLCommons - Aishwarya Srinivasan, AI at Fireworks AI - Subha Tatavarti, Advisor and…
The cloud isn’t the future of AI. BlackBird 1.0 is here — build smart, autonomous agents that run entirely on your device. Download now: decompute.run No cloud. No streaming. Just private, local intelligence. AI that’s fast, sovereign, and truly yours. Welcome to the…
I have been asked, why do we at @DecomputeAI care about better and efficient fine-tuning if one can do RAG? As a mathematician, precision and logical depth are non-negotiable for me. So, I tested @DecomputeAI's BlackBird with Calculus Without Derivatives, a graduate-level book.…
We are at HumanX conference in Las Vegas! Meet our team including cofounders @hinadixit & @jalajupadhyay . Come play with BlackBird, an agentive platform based on our breakthrough technology LaserTune with the lowest memory footprint for fine-tuning models in the world. (We take…
The power of intelligence, in your hands. @DecomputeAI Blackbird gives you answers and keeps your secrets on your device. Your personalized AI can help you write, research, analyze and build your code – you name it.
🚀 Meet LaserTune: the cutting-edge fine-tuning algorithm that’s 2x faster & slashes space usage by 50% compared to DeepSeed! Powering @DecomputeAI's Blackbird, it also shows the “catapult” effect in training loss, signaling better generalization. decompute.run/post/introduci…
