Skip to content

Projects

Things I've built, sorted into models, the systems that run them, and tools for other builders.

16 projects
QuestionTriageDocsWebToolsQuery triage
SystemsMulti-agent knowledge assistant
Agents that route each question, retrieve from two knowledge sources and check the answer before replying.
Case study
WavelengthHundreds of bands per pixel
ModelsCNN and ViT for hyperspectral imagery
A hybrid convolution and transformer model that maps land cover from hyperspectral images with very few labels.
Read the paper
> design a model for flood change detectionTask: change detection · Sentinel-1 SAR + S2Proposing: Siamese U-Net, ResNet-34 encoderConfigured 3 candidates, benchmarking…Agent skill
Toolsrs-dl-architect
An open-source skill that gives coding agents expertise in designing vision models for satellite imagery.
View on GitHub

All projects

  • Engineered a LangChain and Google Gemini powered multi-agent system with intelligent query triage, dual-source RAG (Parent-Child retrieval, query expansion, Flashrank re-ranking), dynamic tool-use for internal/external data access, and guardrail validation for reliable, domain-specific query handling.
    • LangChain
    • Gemini
    • FAISS
    • RAG
    View on GitHub ↗
  • Hyperspectral sensors record hundreds of narrow bands for every pixel, but labelled training data is scarce. This work compares 3D-CNNs, graph networks, CNNs and vision transformers on that problem, then pairs a Spectralformer backbone with SSRN-style spectral-spatial preprocessing.

    Trained on only 10% of the labels, the model reaches 99.80% accuracy on Indian Pines and 97.54% on Pavia University. The work is published in Springer LNCS (PReMI).

    • PyTorch
    • Vision Transformer
    • Hyperspectral
    View on GitHub ↗
  • rs-dl-architect is a lightweight, end-to-end framework built to design and evaluate deep learning architectures for remote sensing and Earth observation data.

    • Agentic automation: Packaged as a reusable agent skill that proposes, configures, and benchmarks model architectures.
    • Targeted tasks: Streamlines semantic segmentation, object detection, and change detection on satellite imagery.
    • Advanced capabilities: Core support for multi-source data fusion (optical/SAR) and multi-temporal analysis.
    • Production-ready: Built-in tooling for benchmarking model performance and ensuring interpretability.
    • Python
    • PyTorch
    • Agent skill
    • Remote sensing
    View on GitHub ↗
  • A multi-channel quantum convolutional network built with TensorFlow Quantum and Cirq, using custom quantum convolution layers and MCQI image encoding. Spectral-intensity stratified sampling balances slum and non-slum training patches.

    The model reaches 94.8% test accuracy and an F1 score of 0.933, ahead of a comparable classical CNN. Published in Quantum Information Processing.

    • TensorFlow Quantum
    • Cirq
    • Sentinel-2
    View on GitHub ↗
  • A PyTorch 3D convolutional network that treats the spectral bands of Sentinel-2 imagery as a third dimension, so spectral and spatial patterns are learned together. It predicts patch by patch and stitches the results back into a full map of slum boundaries.

    • PyTorch
    • 3D CNN
    • Sentinel-2
  • U-Net, U-Net++ with an EfficientNet-B0 encoder, and FPN models trained on the VinDr-RibCXR dataset to segment and label each of the 20 ribs in a chest X-ray. The best model reaches a Dice score of 0.834 and holds up on the separate JSRT dataset.

    • PyTorch
    • U-Net++
    • FPN
    • Medical imaging
    View on GitHub ↗
  • A compact residual CNN that upscales game frames four times, from 480×270 to 1920×1080. Eight residual blocks and PixelShuffle upsampling keep the model small, and mixed-precision training on 4,500 image pairs keeps it fast.

    It beats bicubic upscaling on PSNR, SSIM and a perceptual quality score.

    • PyTorch
    • CNN
    • Super-resolution
    View on GitHub ↗
  • Combines YOLOv5 detection, DeepSORT tracking and a ResNet-50 re-identification model, so a person keeps the same ID as they move between cameras. It re-identifies people with 87% accuracy, runs in real time on standard hardware and is served through FastAPI.

    • YOLOv5
    • DeepSORT
    • FastAPI
    View on GitHub ↗
  • FinBERT, BERT and DistilBERT fine-tuned on 4,840 expert-labelled sentences from the Financial PhraseBank and compared side by side. FinBERT comes out ahead with 90.9% accuracy and a weighted F1 of 0.91, against 0.86 for BERT and 0.81 for DistilBERT.

    An optimizer and learning-rate study shows how each model converges.

    • Hugging Face
    • FinBERT
    • PyTorch
    View on GitHub ↗
  • A serverless React and Firebase app for managing projects, tasks, habits and a calendar in one place. Gemini breaks a goal into concrete tasks, and Google Calendar shows them in your schedule through client-side OAuth.

    Firestore security rules keep each person's data private.

    • React
    • Firebase
    • Gemini API
    View on GitHub ↗
  • A Siamese LSTM with pre-trained GloVe embeddings, trained on the Microsoft Research Paraphrase Corpus. It reaches 91.4% test accuracy, with every experiment tracked in MLflow.

    • TensorFlow
    • LSTM
    • MLflow
    View on GitHub ↗
  • Checks stationarity and seasonality with ADF tests, ACF/PACF plots and decomposition, then compares ARIMA, SARIMA and LSTM forecasts. A stateful LSTM gives the best fit, with an RMSE of 10.3 and R² of 0.83.

    • Statsmodels
    • LSTM
    • Pandas
    View on GitHub ↗
  • A disaster-preparedness web app with a Leaflet map, live geolocation, filters and flood-zone layers from PostGIS. A FastAPI backend with GiST indexes and Redis caching makes location queries 99% faster. Deployed on Google Cloud.

    • FastAPI
    • PostGIS
    • Redis
    • Leaflet
    View on GitHub ↗
  • A browser dashboard built on OpenLayers that connects to GeoServer through the OGC WMS and WFS standards. You can browse layers, query features and run simple analyses on data stored in PostGIS.

    • OpenLayers
    • GeoServer
    • PostGIS
    View on GitHub ↗
  • A lightweight client for OGC Sensor Observation Service endpoints such as istSOS. It discovers sensors, places them on a map and plots their readings as tables and time-series charts.

    • JavaScript
    • istSOS
    • Google Charts
    View on GitHub ↗
  • A reimplementation of the Triple Generative Adversarial Nets paper that generates labelled synthetic images, reaching 99.10% peak classification accuracy. It also compares VAEs and autoregressive models as other ways to make synthetic training data.

    • GANs
    • Semi-supervised
    View on GitHub ↗