Selected work
Projects.
Research and applied work across computer vision, LLMs, agentic AI, NLP, program synthesis, convex optimization, DSP, generative models, and molecular ML. Click any card for details.
Computer Vision
Understanding what neural nets see, and tracking things through space and time.
tf_cnnvis
A library for visualizing feature learning inside convolutional neural networks. Poster at NVIDIA GTC EU 2018.
Food Detection — on-device nutrition estimation
Segmentation + multi-view volume estimation for a food/nutrition client. Shipped 2022–2024 across web and mobile.
Real-time plastic sorting on a conveyor belt
Detecting and classifying moving plastic items in under 100 ms via TensorRT-optimized detectors with in-graph NMS. 2023–2024.
Semantic search over a curated art collection
CLIP-embedding vector search + Elasticsearch metadata filters for an art-curation client. Delivered Jan–Mar 2024.
PoseSync
Robust pose-based video synchronization for sports — pose detection plus graph-based correspondence matching.
Faster real-time tracking
Pedestrian tracking pipeline combining a YOLOv3 detector with Kalman filtering for identity persistence.
Tiny object detection
Detection pipelines tuned for very small objects in high-resolution aerial imagery.
Specularity removal
Robust road detection under specular highlights. Published at IEEE TENCON 2016.
LLMs
Retrieval, reasoning, and generation on top of large language models.
Customer survey analysis with LLMs
Automating UX research at scale — fine-tuned question classification, traceable summaries, an AI interviewer, and mission-level RAG Q&A. Led 2023–2026.
Scientific content summarization & blogification
End-to-end GenAI marketing pipeline for a microhabit-tracking platform — research-grounded blogs, Midjourney art, SEO, approvals, and multi-channel fan-out. 2023–2025.
RAG-based QA & summarization
Retrieval-augmented question answering, summarization, and content-generation systems built on hosted and open LLMs.
Agentic AI
Systems that plan, call tools, and act on natural-language intent.
Hazard Forecasting & Property Intelligence Platform
Architect on an agentic AI platform for insurance underwriters — built on Google's ADK, DeepMind, and Google Earth Engine data.
NL → action-set generation
Automatic generation of executable action sets in response to natural-language requests. US patent application 18/142,472.
NLP
Sequence modelling, before and after transformers.
Reasoning & Program Synthesis
Learning programs and abstractions from very few examples.
Convex Optimization
Turning messy domain constraints into problems solvers can chew on.
Algorithmic trading strategy discovery
Cerebellum Capital engagement (2014–2019) — GP-based strategy search, a realistic convex portfolio optimizer, and time-series forecasting across 2,000+ stocks.
Robust supply chain planning for a global apparel retailer
Digital twin + network-wide convex optimizer for ~20 plants, 30+ DCs, thousands of stores. Architect + convex formulation. 2020–2022.
Digital Signal Processing
Signals — measured, mapped across modalities, and tracked under noise.
Kalman-filter GPS receiver
ISRO internship — a GPS receiver simulation robust to ionospheric scintillations, using a Kalman-filter based tracking loop.
EEG → audio mapping
Cross-modal representations that map EEG signals to audio signals — exploring the shape of brain-activity embeddings.
Generative Models
Learning distributions well enough to sample from them.
Molecular & Bio ML
Neural representations of molecules and cells, tuned toward useful properties.
Directional VAE for molecules
Continuous molecular embeddings via directional variational transformers; property prediction & discovery with VAE + Bayesian optimization.
scaLR — low-resource single-cell analysis
End-to-end deep-learning platform for scRNA-seq analysis and biomarker discovery on modest hardware. Briefings in Bioinformatics, 2025.
Depolymerization catalyst discovery
ML platform to find solid catalysts for depolymerization reactions. US patent application 18/435,957.