AutoML system that intakes arbitrary rectangular data and builds and compares thousands of deep neural networks via genetic algorithms (DeepNEAT) to build the optimal model for your data. This system is highly performant on both flat and time series data, having the capacity to encode many variations of RNNs and CNNs in addition to standard dense networks.
Leveraged Genetic Algorithms to schedule optimal task orderings across all construction staff and projects. This system solved a more complicated version of the Job Shop Scheduling Problem, an NP-Hard problem that is “clearly harder than the Traveling Salesman problem.” The system optimized over individual staff skillsets, drive time, site visit ordering (traveling salesman), task ordering, resource availability, wage spend, and more.
This system was built by adapting multiple bleeding edge research papers into a full production AI system. The system accepts streams of realtime flight data, and thus has extremely low latency requirements. It then performs unsupervised clustering of timeseries data in order to flag potentially dangerous flight paths.
This time recommendation engine takes an arbitrary number of calendar schedules as input and understands the context of the event based on features such as NLP embeddings of event name and description, as well as history of users past events. It output times that worked well for everyone and also made the most sense for the event in question.
Our location recommender utilized natural language embeddings of massive databases of location details to match locations to groups of users and their respective social events. This engine allowed for curated social events to be built automatically with everyone’s preferences in mind.
Computer vision model that takes in an image of a receipt and outputs an itemization of all charges, each tagged with relevant item categories. To make this work properly, we needed to create a custom pipeline comprised of many deep neural networks. The first model performed object recognition, passed the output to an OCR model, then a model that we trained to understand receipt geometry, and a finally an NLP model that performed classification on the types of charges.
This system intakes 3 view mammogram scans of breast tissue and outputs the risk factors associated with the images. It improves on the current clinical accuracy of risk assessment by more than 10% in the majority demographic (caucasian women) and up to 60% improvement in the case of minority demographics. We adapted just released research into a production grade AI system and helped oversee and advise the process for technology acceptance by the FDA.
Utilized ASR to transcribe 20+ years of cataloged audio records from a multimedia company, then created a semantic search engine via natural language embeddings to retrieve relevant recordings and pinpoint the most relevant audio segments down to the second level.
Trained a hyperspecific Named Entity Recognition (NER) and Entity Linking model on hundreds of thousands of industry specific documents which resulted in near full automation of multiple job types in the food regulatory industry, where inaccurate predictions have severe consequences. Coupled this with a powerful graphical interface for verifying predictions downstream to produce a production application on Azure Cloud.
Trained a model to reliably extract event details such as title, date, time, participants, and location from blocks of unstructured text and automatically create standardized event entries. This enabled event extraction from both the internet and from conversations, as well as instant event creation via voice.