Leo Pekelis is an applied statistician and AI research leader who helped develop always-valid statistical inference, specialized financial language models, and open models with million-token context windows. Formerly chief scientist at Gradient, later DeepSky, he joined Airtable when it acquired the company.
Pekelis earned a bachelor’s degree in economics and a master’s degree and doctorate in statistics from Stanford University. At Optimizely, he helped create Optimizely Stats Engine, combining sequential testing with false-discovery controls so teams could repeatedly inspect A/B tests without undermining their conclusions. His research with Ramesh Johari and David Walsh developed p-values and confidence intervals that remain valid as continuously monitored experiments evolve.
After roles at Opendoor and Disney, Pekelis became CloudTrucks’ first data hire and ultimately its head of data, building teams and systems spanning analytics, infrastructure, machine learning, freight recommendations, routing, scheduling, and pricing. He also advised experimentation company Eppo before joining Gradient as chief scientist in 2024.
- Finance-specific model training. Pekelis helped develop Albatross, a Llama 2-based financial language model trained through automated document screening, specialist review, synthetic augmentation, continual pretraining, supervised fine-tuning, and preference optimization. He distinguishes between teaching a model financial knowledge and training it to apply that knowledge to calculations, financial tables, summaries, and grounded answers.
- Long-context grounding. Pekelis argues that inserting relevant documents and examples directly into a model’s context can reduce hallucinations and avoid retrieval failures when evidence depends on relationships across multiple passages. His examples include prompts containing thousands of demonstrations and experiments using anonymized books to capture an author’s style and themes.
Following Airtable’s acquisition of DeepSky, Pekelis has focused on statistically robust AI systems for research agents and business workflows: systems that remain useful when organizational data is noisy, prompts are imperfect, and relevant context is extensive.
