Building architectures that
learn without forgetting.
My current research is focused on continual learning and trustworthy AI — designing model architectures that can keep learning new things without erasing what they already know, and studying how to intervene on what a model has learned rather than only what it says. I am the author of TFGN and the founder of VedaBio.
I am currently doing independent research in machine learning. My career has had one throughline across very different substrates: take a mechanism, engineer it into a working system, and validate it rigorously enough to ship. My first job out of IIT was warehouse optimization and growing the sports category at Flipkart (acquired by Walmart). I then moved into biotechnology — a PhD in bioengineering at the University of Illinois Urbana-Champaign, where I co-created foundational technologies for cancer research and infectious-disease diagnostics (cumulatively 50+ patents and 30+ publications) — and spun that work out as VedaBio (formerly LabSimply; Y Combinator S21), which I led as Founder/CEO of a 30+ person team through a $40M+ Series A.
In 2025 I turned toward software and built AI-native applications, including WorkManga (more below). Most of that work lived in the application layer — largely orchestrating a base model — and it gave me little real control: I couldn't explain why the technology broke the way it did, which pulled me down into the architectural layer. Reading the literature, I gravitated to continual learning; a decade of looking at biology as an engineer — and at how memory actually works in biological systems — gave me a way into it that became TFGN.
My research centers on continual learning — how models keep learning without erasing what they know — approached through how memory works in biological systems. TFGN is the first result of that work.
TFGN: Task-Free, Replay-Free Continual Pre-Training Without Catastrophic Forgetting at LLM Scale
A transformer overlay that keeps the forward pass dense and shared while structuring cross-domain parameter updates into near-orthogonal subspaces — a structural gradient-orthogonality signature (≥99.59% L2-orthogonal) with operator-level reshape demonstrated at ~0.9996 cosine fidelity. Sole author; all experiments from ~398M to ~9B parameters.
Write-side interventions for trustworthy AI
Whether reshaping what a model effectively computes — rather than steering activations — yields interventions that generalize under distribution shift, compose, and resist adversarial removal.
Continual learning without catastrophic forgetting
Architectures that keep learning across domains and tasks with no replay buffer and no task boundaries, while preserving prior capability.
A decade of mechanism-to-system research
Spanning graphene transistors and biological sensors of many kinds, through chemistry and assay design, to CRISPR molecular detection (VedaBio) and spatial transcriptomics — with mechanical-engineering origins. ~2,000 citations; 50+ patents.
Selected from ~30 peer-reviewed papers, ordered by journal prominence. Impact factors: Clarivate Journal Citation Reports (most recent release).
In 2025 I moved into software and shipped the projects below — the experience that pulled me toward the research above.
Tribl Shopping
Shipped · now inactiveMy 0→1 software experience — a community-led alternative to high-markup grocery delivery.
Motivation: I hated hidden markups in delivery apps. Tribl flipped the model — transparent in-store pricing, neighbors form "tribes," and everyone splits the real bill. I designed and built a React-Native grocery app (iOS / Android) and a companion driver app ("Tribl Drivers"), and served customers in San Diego.
I ran every growth loop and channel myself: door-to-door, Meta/IG ads, AI-generated creative and UGC, and influencer pilots (a surprising breakout segment — women 55+). The lesson wasn't technical: marketing and go-to-market are disciplines in their own right, taking a different mindset and a different kind of perseverance than building the product. Demand turned out to be lukewarm, and beyond the learning I didn't find it technically rewarding — so I pivoted.
WorkManga
Shipped · now inactiveMy second project and first AI-native application — a voice-first operating system for work.
A voice-first, contextually-aware, self-organizing file manager and unified inbox for all your work — think, make decisions, and take actions with zero context switch. It unified email, messages, and documents, auto-triaged them into projects, drafted memory-aware responses, and let you call or text an agent to get updates and act by voice. I built it as the tool I wished I'd had when I was leading a 30+ person team.
Key learning: the friction of getting a user to connect anything beyond email was the wall. Without that, the product collapses back into being just another email client — which is where the real product-design problem lives.
Voice AI for trucking
DemoA voice system that truck operators can use.
We built a working voice-AI demo aimed at trucking operations. The most valuable part was a couple of weeks spent getting deep into how trucking actually works on the ground — the kind of domain immersion that changes how you design a product.
Working through these is what surfaced the questions that became the research above.
Notes and essays on my research — intuitions and explanations meant to be readable.
Open to research conversations, collaborations, and questions about TFGN or the trustworthy-AI work.