Research contributor, LEMS, Brown University Master's student in Applied Artificial Intelligence, Northeastern University
Computer vision: perception, 3D geometry and visual-inertial state estimation, for autonomous systems and computational biology.
I build vision systems that perceive, localize and act under uncertainty. At Brown, in Prof. Benjamin Kimia's group, I work on GPS-denied drone navigation and own the synchronized multi-camera sensing layer its visual pipeline depends on. At Northeastern I develop deep-learning detection and visual–IMU sensor fusion for the NUWave underwater ROV. I carry the same perception-to-decision loop into computational biology. I am applying to CS PhD programs for Fall 2027, after five years as a software engineer.
My research aims to make visual perception reliable enough to act on: systems that recover geometry and semantics from imperfect sensors, generate plausible structure where data is missing, and know how far to trust what they produce.
Visual perception and 3D vision
Recovering structure and meaning from images: multi-camera geometry and calibration, learned detection and recognition, and representations that hold up under blur, low texture and domain shift.
Multiview geometry, camera models, CNN detectors, diffusion models, representation learning
State estimation for autonomy
Fusing vision with inertial and other cues so aerial and underwater vehicles stay localized where GPS fails and visual features are scarce.
The diffusion and 3D reconstruction machinery of computer vision, turned on molecules: generating structures under learned constraints, recovering them from cryo-EM images, and deciding which predictions are trustworthy enough to test.
Diffusion and flow matching, classifier guidance, cryo-EM reconstruction, calibrated active learning
News
Preliminary negative result: calibrating the scoring model does not make closed-loop molecular discovery more efficient. Read the research note.
Joined Prof. Benjamin Kimia's group at Brown (LEMS) on GPS-denied drone navigation, starting with the multi-camera sensing layer.
Joined Northeastern Marine Robotics, working on deep-learning detection and visual–IMU fusion for the NUWave ROV.
Research
GPS-Denied Visual Navigation for Drones
Prof. Benjamin Kimia's group, LEMS, Brown University Research contributor, 2026 to present
Ongoing
Localizing a fast-moving drone without satellite navigation, where motion blur and low texture starve any single view of features. The approach integrates many synchronized views with inertial cues, so every downstream estimate inherits the geometry and timing of the camera rig.
My part. I own the multi-camera sensing layer the visual pipeline rests on: hardware bring-up of a synchronized camera array on a Raspberry Pi 5, camera calibration, and frame-level time synchronization, with calibration and test code contributed to the group's repository.
Multi-camera systems, camera calibration, time synchronization, visual–inertial navigation
NUWave: Deep Perception and Sensor Fusion Underwater
Team research project, Northeastern Marine Robotics MATE ROV competition team, 2026 to present
Ongoing
Underwater perception for a semi-autonomous ROV, where turbidity, refraction and changing light break assumptions that hold in air.
My part. I develop deep-learning object detection and visual–IMU sensor fusion in the ROV's ROS perception stack, including the recognition node and its automated tests.
Deep object detection, visual–IMU fusion, ROS, OpenCV, I²C and UART
Calibration-Aware Active Learning for Closed-Loop Discovery
Independent research, self-directed Bridges autonomous decision-making and computational biology
Preliminary result
A self-driving lab is a control loop: a model proposes candidates, a system tests a subset, and the results retrain the model. I simulated that loop over 2,311 PTP1B inhibitors from ChEMBL and asked whether calibrating the scoring model makes it more efficient.
Result. It does not: 35.5 ± 3.5 hits versus 35.3 ± 4.2 at a 360-experiment budget, and for single-objective selection it hurts. The mechanism is resolution loss: isotonic regression is monotone but piecewise constant, so it collapses 1,844 distinct scores into 74 and leaves 100 molecules tied at exactly 1.000, making batch selection arbitrary among them. One target and 8 seeds so far.
Active learning, calibration, sequential decisions, ChEMBL, RDKit
Classifier-Guided Diffusion for Post-Stroke Drug Design
Course project, Applied AI (AAI 6600), Northeastern University
In progress
Most stroke survivors miss the treatment window and keep lasting disability, partly because scar tissue blocks nerve regrowth. The pipeline generates candidate molecules with a diffusion model and steers every denoising step with the gradients of property classifiers for blood–brain-barrier crossing, target activity and toxicity, the same guidance mechanism used to condition image diffusion.
Design. Guidance classifiers are trained on noised inputs, and a separate held-out classifier judges the final candidates, so the generator cannot succeed by exploiting its own guide. All scorers are validated on scaffold splits, so reported performance reflects new chemistry rather than memorized analogs.
Course project, Applied Deep Learning (AAI 6640), Northeastern University Team lead of four
In progress
A recurrent network that forecasts a full probability distribution rather than a point estimate, so any threshold question is read straight off the model. It extends the GARCH–LSTM equivalence of Zhao et al. (2024) with an in-cell variance recursion and a Student-t output head.
My part. I lead the team and own the ten-year dataset and the evaluation: purged walk-forward splits, isotonic recalibration and conformal coverage.
What I would like to work on in a PhD, each one grown out of the work above.
How should uncertainty in a perception system's geometry, from calibration to pose to depth, propagate into the decisions built on it?
Multi-camera estimates at Brown and score calibration in closed-loop discovery share a failure mode: small, invisible errors become confident downstream choices.
How can visual–inertial estimation stay reliable when fast motion and low texture starve the visual front end?
The GPS-denied drone and the underwater ROV meet this from opposite ends: blur at speed in open air, and murky water with little to track.
Can the diffusion and 3D reconstruction tools of computer vision recover and design biological structure, and know when their outputs deserve an experiment?
Cryo-EM is 3D reconstruction from projections with unknown pose, and AlphaFold 3 and RFdiffusion are diffusion over coordinates instead of pixels. My closed-loop result suggests the last step, deciding which outputs to trust, deserves as much attention as generation itself.
Methods and tools
Vision and sensing
Camera calibration, multi-camera synchronization, CNN-based object detection, OpenCV, embedded vision on Raspberry Pi 5
Estimation and robotics
Visual–IMU sensor fusion, ROS perception stacks, embedded camera hardware bring-up
Learning
PyTorch, TensorFlow and TensorFlow Probability, sequence models, reinforcement learning (Stable-Baselines3, Isaac Gym, MuJoCo)
Uncertainty and evaluation
Isotonic recalibration, conformal prediction, active learning, scaffold-split and purged walk-forward validation, ChEMBL and RDKit
CAD-LLM, which turns natural-language part specifications into structured CAD parameters, and APML, a predictive pipeline that benchmarks models on both accuracy and inference latency.
Chess
Member of the Northeastern University Chess Club.
Contact
I'm glad to hear about research collaborations and PhD openings in computer vision, 3D perception, autonomous systems and computational biology.
