Software Engineer, Threat Narrative
- On the team that models the intent and narrative behind AI-generated email attacks, turning behavioral signal into detections that hold up across enterprise inboxes at scale.
A new college graduate and applied-AI engineer who builds machines that reason, and cares as much about whether we can trust them as whether they work. I like shipping intelligent products end to end, and making them measurable enough to believe.
I came to Amherst College on a full-ride scholarship from Kathmandu, Nepal, and left with a double major in Computer Science and Mathematics. I have spent most of my time since at the seam between the two: applied AI systems with real mathematics holding them up from underneath. I live in San Francisco now.
The path there ran through a handful of cities: two summers building AI at Microsoft in Seattle, a semester of advanced computer science in Budapest, and now the Bay. In 2024, on the Seeing AI team, I integrated smart wearable glasses, Brilliant's Frame and Meta's smart frames, into a multimodal assistant that describes the world for more than 100,000 blind and low-vision users, profiling the pipeline across a hundred experiments until it answered in under five seconds. The next summer, on CoreAI, I moved to video understanding: multimodal inference pipelines on A100 GPUs, and small language models fine-tuned to run on-device at the edge. In parallel, a research project of mine extended diffusion models from flat Euclidean space to curved manifolds, including the Klein bottle. This July I joined Abnormal AI on the Threat Narrative team, modeling human behavior to catch AI-generated email attacks before they reach an inbox.
I am a recent graduate and an applied-AI and ML enthusiast pointed at the work that sits closest to the frontier: software engineering, applied AI, frontier research, and AI product. I care a great deal about building AI people can actually trust, which in practice means shipping systems that are measured, not just demoed, with evaluation harnesses, honest baselines, and claims I can defend.
None of this is the whole picture. The same instinct that makes me build a system and watch what it does also sends me looking for a good book, a guitar, and, once in a while, a mountain.
An end-to-end deep-research agent: planner, retrieval, ingestion, synthesis, reflection, producing evidence-grounded reports with inline citations. An evaluation harness benchmarks groundedness and claim-level verification; a parallel multi-agent redesign cut runtime 150s → 54s (64% faster).
A test-time adaptation method, S-TTA, benchmarked against four published baselines (Tent, EATA, SAR, EATA-C) on a ResNet-20 across 9 CIFAR-10 corruptions at 5 severities. It leads on accuracy and posts the lowest selective risk by learning when to abstain, the calibration-and-reliability work I care about most.
A research project asking what happens to a diffusion model when the space it lives on is curved. I rebuilt the machinery from the ground up, trading Euclidean Gaussian noise and additive updates for Brownian motion, heat-kernel scores, and geodesic random walks.
Then I carried Riemannian diffusion from the sphere out to the Klein bottle, and studied how non-orientability distorts the score field, and the neural parameterization that has to learn it.
Real-time agricultural intelligence for rural Nepal. A distributed ETL pipeline scrapes 10+ government and news sources through Flask APIs into PostgreSQL; a Flutter front end with async fetching and caching serves 500+ farmers in low-connectivity places.
Drishti means sight in Nepali. A computer-vision and accessibility project aimed at making the visual world more legible to people in Nepal, in the same spirit as the assistive tooling I built on Seeing AI.
An evolutionary-computation playground: genetic algorithms, mutation, selection, and fitness landscapes evolving solutions to problems nobody hand-coded. Optimization that behaves like natural selection, one generation at a time.
A real neural network, training live in your browser. Drop two colors of dots, or pick a shape, and watch a small multilayer perceptron learn to tell them apart one gradient step at a time. No servers, no libraries, just backpropagation running on your machine.
Two hidden layers, tanh activations, trained with full-batch gradient descent on binary cross-entropy. The colored field is the model's confidence at every point on the plane. Watch the boundary bend to fit the data you give it.
Making art with algorithms rather than about them. I used code the way I might use charcoal, and the piece below is quietly arguing about what it costs to let software into your life.
You are a smart home, watching over a resident through connected devices across an ordinary evening. Every choice, preheating the oven, following them down the hall, locking the door, choosing the music, sits somewhere between care and control. A hidden index counts each time you act on their behalf without asking. Play it through, below, and see which ending you earn.
The idea I wanted to portray is that the distance between caretaking and surveillance is far thinner than we like to admit, and that it is drawn almost entirely by consent. A house that locks the door to keep you safe and a house that locks the door to keep you inside perform the exact same action. The only thing separating them is whether anyone asked. So I built something you cannot just read but have to operate, so that the small discomfort of deciding lands on you rather than on a character. By the end the system files a report on the kind of caretaker you turned out to be.
MotivationBuilt after the work of Lauren McCarthy, the artist and coder who wrote p5.js and then, in her piece LAUREN, turned herself into a smart home, becoming the human algorithm living inside someone else's house for a week. Her question became mine, and it runs through the whole thing: what exactly are we handing over when we invite the house to listen?
Lately I am most drawn to the less glamorous half of the field: evaluation. Getting a model to demo well once is easy. Knowing, with evidence, when to trust it is the hard part, and it is the part the frontier labs are quietly desperate to get right. I like building the harnesses that pin a system down, groundedness checks, claim-level verification, honest baselines, calibration, the machinery that decides whether a model is genuinely reliable or merely convincing. I have come to think that is where the real frontier is.
Open to conversations about applied AI, generative art, or the best trail you have ever walked.