Constantin Weisser

Haize Labs

Employee #1 and Member of Technical Staff. Working on dynamic evaluations for AI safety. e.g. Active Learning for Aligning LLM judges to human evaluations

ML Alignment and Theory Scholars

Demonstrated that targeted manipulation and deception emerge in LLMs trained on user rather than annotator feedback. oral @ SATA & spotlight @ SoLaR workshops NeurIPS 2024. [MATS 6.0 supervisor: Micah Carroll]

NASA Frontier Development Lab

Conceptualized and ran the McKinsey partnership with FDL for a 9-week research sprint to tackle climate change through the world's first accurate real-time induced seismicity forecasting system. Ran a team of four PhD-level interdisciplinary researchers plus ~10 supportes as ML and management lead. By reducing the runtime of SOTA approaches from 22 hours with manual setup to 2 minutes without, our team addressed a main pain point of the seismicity community and fulfilled a central goal of the DoE's multi-institution and -year SMART initiative.

QuantumBlack, a McKinsey company

Led the development and implementation of cutting-edge AI solutions for Fortune 500 clients' most critical business challenges. Spearheaded multiple cross-functional teams across organizations to analyze requirements, architect technical solutions, and rapidly deliver production-ready systems with measurable business impact. Successfully delivered machine learning solutions across six different industries, including leading the end-to-end development of an enterprise-grade conversational AI system with robust evaluation frameworks and safety guardrails.

Amazon Alexa Health NLU

In 2020, when the regime of AI foundations model was just starting, I worked at Amazon Alexa to supparize medical text. After solving the business problem by running the SOTA summarization model, I developed an LLM sampling algorithm that removed pathologies and improved fluency.

MIT PhD in Physics, Statistics, and Data Science

Worked on machine learning for particle physics supervised by Mike Williams. At CERN, sifted through 200TB of LHCb data to better understand dark matter, setting world-leading limits on dark photons [163 citations]. Made the following contributions to ML for Science:

1. Built first high-precision generative model approach to high energy physics simulation [ICLR '20]
2. Developed loss to suppress higher-order classifier dependence on a nuisance parameter [NIPS '20]
3. Developed a powerful novel deep learning approach to finding the vertices in particle collisions
4. Developed fast, physics-informed surrogate model for coastal floods via FNOs [NIPS'21]

Technical Competitions

Over the years I have taken part in various technical competitions, which test team work and under pressure:

• 1st place Amazon Nashville Code Competition against 26 SDEs.
• 2nd place Yale Graduate Case Competition against 28 teams.
• Best science case as Head of Science for European Space Agency's Alpbach Summer School.
• 3rd place MIT IAP ML in Critical Care hackathon.