Sean Condon

At learn ventures, I work on a variety of machine learning and physics simulation approaches to design novel proteins that can be used for treating a multitude of medical conditions. To date, our AI-based drug design pipeline has produced 1000s of new proteins which have been synthesized in lab and are currently undergoing testing for effectiveness in a variety of tasks. This work includes reading new papers every day on state-of-the-art ML and simulation based approaches for drug design, re-implementing models in our own infrastructure, and working with biochemical labs to verify that our in-silico methods work in the real world.

pytorch   GCP   CNNs / transformers

The Large Hadron Collider (LHCb) at CERN collides atomic particles, and then probes the remnants of these collisions to make experimental breakthoughs in particle physics. The LHCb detector outputs data at a rate of 40 Tbit / s, and this data must be processed in real-time to determine what is useful and what can be thrown away. This determination was originally done by simply threshholding the collision parameters in a cut-based method. During summer 2020, I trained decision tree classifiers to handle this determination instead, allowing for 15% greater accuracy at the same false positive rate compared to cut-based classifiers. The decision trees were implemented in CUDA so that they could run on LHCb's GPU clusters during round 3 of data collection.

scikit-learn   CUDA  decision trees

Enabled by the internet of things, the weather predictions of tomorrow will come not just from weather stations, but also from smartphone sensors, traffic cameras, and many other internet-enabled devices. Cellphone towers, for example, offer a way to measure real-time precipation where no traditional measurement devices are available, as cell signals attenuate while passing through rainfall. During January 2020, I re-parameterized and re-trained an LSTM model to predict rainfall rates based on cell-tower attenuation, improving on Tomorrow.io's older implementation with a 5% accuracy boost. The new and improved model was deployed to run precipitation predictions on 60,000 cell tower links throughout the United States.

tensorflow   GCP   LSTMs / RNNs

Gravitational waves form when two black holes collide and send out ripples in space-time. These ripples can be measured by very sensitive interferometers like LIGO, where a complex algorithm detects faint signals in exceptionally noisy data. In past years, this algorithm has been a match-filtering approach that compares time-series to an extensive template bank of possible signals. However, during 2019, I demonstrated that convolutional neural networks could effectively relpace match-filtering approaches, offering lower latency and higher abillity to interpolate between templates. My CNN architecture performed 100x faster than the old approach with similar accuracy, allowing a job that previously took 800 CPU cores to run on a single GPU.

tensorflow   GCP   CNNs

Many planets - called exoplanets - exist outside of our solar system. They are first identified as objects of interest by large sky surveys from telescopes like Kepler, and then verified as exoplanets with follow-up radial velocity measurments. During summer of 2018, I characterized 3 new exoplanets from the Kepler object of interest catalog. I combined data from 340 stellar spectra measurements with ground-based radial velocity measurements using the standard Python scientific libraries (numpy, scipy, matplotlib, pandas). I then derived the exoplanet parameters that could explain these spectra / radial velocity measurements using a Markov Chain Monte Carlo algorithm.

scipy / matplotlib / numpy   monte carlo