Research, implementation, and evaluation of AI systems for investor (VC/PE) insight into emerging technological landscapes, including retrieval-augmented generative (RAG) workflows combining our in-house AI platform for patent language modeling developed at Aon) with open source & proprietary foundational LLMs (OpenAI, Anthropic Claude, AWS Bedrock, etc.).Note: Moat Metrics was previously known as Aon IPS (Intellectual Property Solutions). The business was incubated within Aon's New Ventures Group, before being spun out into an independent entity.

Collaborated on the development of a custom transformer-based large language model (LLM) with multi-task fine-tuning. The model embedded texts and labels from patents, products, and markets into memory-efficient 128d dense vectors, accurately predicted applicability to ~8,000 high-granularity imbalanced market classes associated with revenue, and provided high performance for retrieval, ranking, and clustering tasks.Invented and prototyped an intellectual property landscape explorer web application for patents, products, companies, markets, and litigation risks, with state-of-the-art clustering, and dynamic visualization of technology trends over time. Contributed on R&D of later versions with semi-supervised unified clustering/classification.Created a custom auto-scaling ML pipeline on AWS (ECS, ECR, SQS message queue, CloudWatch) empowering analysts to label and train classification models and do full corpus prediction ~5000% faster than the previous SageMaker solution, and enabling an iterative active learning strategy, reducing expensive analyst labeling effort.Continuously monitored ML/AI research developments for applicability to our business needs, guiding the group direction. Example: Discovered a reconfigurable inverted index for approximate nearest neighbor subset search enabled a more accurate, fast and memory-efficient patent prior-art search, and in turn, a patent validity model (predicting the likelihood of a patent surviving a PTAB challenge).Designed, implemented, and evaluated a system to generate “patent-implied” financial metrics (like price-to-earnings, price-to-sales, R&D replication cost, etc.) for companies from sets of point-in-time patent (and/or product) embeddings and SEC data. Scaled the system with AWS Batch. Evaluated these metrics in quantitative investment strategies. Some demonstrated alpha in backtests. Predicted valuation metrics for private companies without publicly available financials.Mentored early-career colleagues.

Provide leadership, instruction, and feedback to a research team working on a senior project in the area of machine learning for real estate and city planning.

Designed, implemented, and maintained both custom software for large clients as well as a new enterprise-grade product called Cloud Migration Manager to ease the onboarding of new customers moving to our cloud-based file-sharing and collaboration system. Data included files, folders, permissions, users and groups. On-premise to cloud was the initial use case. The product was later extended to enable some cloud to cloud migrations.● Shared backend dev responsibilities equally with one other teammate● Took full ownership of all “TrueUp” features where we maintained filesystem snapshots and performed diffs into order to detect changes and built more intelligent, controlled, and auditable migration features on top of this system● Took primary ownership of security features re NTFS permissions● Contributed greatly to UI design and reviewed/debugged front-end code● Communicated directly with large customers (IKEA, Coach, PCL Construction, Boston Consulting Group (BCG), etc.) to help them plan and carry out their migration projectsKeywords: message queues, workers, SMB/CIFS, pysmb, Redis, Flask, React, Docker, GCP, Google Cloud (Datastore, Pub/Sub)

Collaborated with founder to evaluate business ideas as well as identify value propositions, competitors, risks, and ways to extract value. In addition to business plan and strategy refinement, contributed to some prototype design and development on multiple products. Projects included a general purpose numerical optimization library, a real-estate investment tool for the rental market, and a peer-to-peer marketplace

Worked on the backend of a mobile application to give users useful recommendations (mining a plethora of social media input data) for an ambitious early stage startup. Enabled product demos given by the executive team in their quest for capital.

Collaborated with other engineers to develop TrueCall, sold to MetroPCS, Verizon, etc., which geo-located every call, text, and data packet on our customer’s nationwide networks and stored this and other meta-data in a custom database solution that enabled real-time (~7 sec delay) and historical (going back 2 years) queries. My contribution to the project was important in the company’s successful acquisition.● Became the lead developer on the backend a little over a year after starting● Trained and managed more than a dozen new hires● Designed, implemented, tested, maintained high-availability data pipeline modules deployed to customer switches which formed a core component of our upgrade from hourly to real-time performance● System uptime was better than 99.99%, and could recover from downtimes of up to 12 hours with no data loss and causing only a minimal/well-bounded load (processor, memory & storage) on customer machines (Lucent switches running Solaris OS i.e. the central hubs for a metro region)● On top of this base, built out real-time and historical analysis, reporting, and interactive visualization/mapping features to help our customers optimize their networks in real-time, reduce customer churn, target billboard advertising, reduce costs of complying with police and federal authorities, and provide executive summaries/reports/drill-downsKeywords: Sockets/Select, Protocol Buffers, Boost, Qt, Solaris

Designed, implemented, maintained features on AJAX VoIP product. Examples included additional numbers in any area or country code which ring to an existing mobile number, customizable call forwarding and call screening, web-accessible voicemail, and online contact backup.

Developed methods and tools for protein active site prediction w/applications in drug discovery doing graduate/post-grad level work as an undergrad in Prof Kimmen Sjolander's phylogenomics research group after receiving an A+ in her upper-division protein informatics course.

Designed and implemented algorithms using techniques from computer vision to address problems which can occur when post-processing an image to add depth of field effects while working as a researcher in Prof. Brian Barsky's OPTICAL group. This was done before the dawn of the smartphone era, but is now commonly referred to as “portrait mode” in high-end smartphones. I was invited to join the group as an undergraduate after my photo-realistic liquid animation final project won the award for highest technical achievement in Prof. Barsky’s upper-division computer graphics course.● Originated the key idea that computer vision/machine learning could provide the needed solutions to the occlusion and discretization problems inherent in this application● Prof. Barsky provided positive feedback and asked me to immediately implement the idea and also write up a paper and presentation on it. I did so independently initially.● We then revised the publication(s) as a 4 person team. I led the team in organizing and directed the work of other undergraduates who contributed additional images to the publication(s)

Created an early prototype (both front and backend) of an automatic transmission design system for engineers at the Ford Motor Company which modeled and predicted the effects of parameter variation on composite error for the reduction of noise (a.k.a "gear whine"). The designer specified their transmission configuration through a web-based UI via canonical graphs. Elements such as gears, clutches, and brakes were nodes and the relationships among them were edges. The backend solved for speeds and torques, predicted transmission errors for individual meshes, and combined them into a composite. A conference paper was later published (Fall 2004) acknowledging my contribution to the project.

Developed secure distributed messaging tools which allowed researchers at different locations to collaborate on experiments as if they were in the same laboratory