Southbridge AI

Southbridge AI builds AI agents to make fragmented organizational data usable: understanding datasets, reconciling sources and transforming information into the formats a task requires. Founded in 2025 by Minu Choi and Hrishi Olickel, the company works with teams in healthcare, finance, genomics and insurance. Its domain-specific AI programs, called hanks, encode workflows such as interpreting healthcare codebooks and reconciling financial data.

Its open-source runtime, Hankweave, organizes those programs into sequences of codons: reusable agent blocks containing prompts, workspaces and behavior checks. Each block receives explicitly selected context, helping keep long runs coherent without carrying every earlier interaction forward. Checkpoints, rollbacks and structured event logs let developers inspect failures and resume work. Repairing a codon carries the fix wherever that block is reused, making maintenance part of the program’s design.

That architecture grew out of a change in Southbridge’s approach. The team abandoned the idea of a single intermediate format for all data transformations after finding that the destination determines too much of the required structure. It shifted toward composing agents into inspectable, repairable sequences. In its February 2026 retrospective, Southbridge described a partner project that catalogued 1,415 health-data files spanning 24 years in four unsupervised hours, identifying inconsistencies and producing a usable codebook. The example connects its runtime work to its underlying purpose: reducing the preparation needed before people can use complex data.