Southbridge... is a company founded by Minu Choi... and Hrishi Olickel... in 2025, with the conviction that we can unlock the world’s data at scale - now that we have general intelligence as a commodity.

All of the world’s data - both yours and your company’s - is stuck behind ageing schemas, hostile engineering designed to keep things from connecting, and sheer data debt that accumulates every day, as we put away more and more data that could help us make better decisions.

We never thought this would be easy.... LLMs are the final unlock in solving data engineering at scale, but they can only see a truly tiny part of the information at any time. So much remains to be done. To get here, we had to:

1. Understand how models see data
2. Coax reliability from LLMs over minutes, then hours, then days
3. Figure out how to treat data and cost as resources to enforce at scale
4. Build a runtime that supports 100x the complexity of today's harnesses, by doing less and doing it well
5. Layer in true persistence to allow agents to move anything anywhere
6. Deploy the first artificial data engineer, armed with a true registry of AI programs and an agentic backplane for long-horizon execution
7. Solve ingest, horizontally, unsupervised

Along the way we’ve been fortunate to work with teams in healthcare, finance, genomics, insurance, and more. We’ve built AI programs - we call them hanks, the first of their kind - for different workflows. A financial exchange shipped a hank to production three weeks after first hearing about us. A research partner watched a year of staff time be saved in four hours.

We believe we will forever be neuron-limited if we can’t share our work and build together. Our runtime for long-horizon execution, hankweave, which we’ve used in production since June 2025, and can operate all major agent harnesses with a unified, cost-aware interface, is open-source. Our experiments, results, and tests are shared as we can.

Hankweave - the code, the hanks, the docs • Antibrittle Agents • Why we started

The easiest demonstration of our work is our first frognu comic. Made with Sonnet 4.5, Nano Banana 1 and Haiku, taking just some information and creating a story, rich characters, plots, and images with extreme character consistency, using fresh eyes to review and fix problems, emotional moments and so on. Hanks mean anyone can now have the same outcome - check out the code here.

←01 / 16→

        ideas · writing · style references
                       │
                       ▼
┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐
│ read +  │→│diversity│→│visualize│→│ select  │→│  story  │
│ explore │ │ review  │ │ + samp  │ │ + spec  │ │ review  │
│  Opus   │ │  Haiku  │ │ Opus+N  │ │  Opus   │ │  Opus   │
└─────────┘ └─────────┘ └─────────┘ └─────────┘ └────┬────┘
                                                     │
       ┌─────────────────────────────────────────────┘
       ▼
╔════════════════════════════════╗
║   generate + review loop  ×3   ║◀  ◐ narrator sentinel  (Haiku)
║     Nano panels, Opus tunes    ║
╚═══════════════╤════════════════╝
                │
                ▼
       ┌─────────┐ ┌─────────┐
       │  cold   │→│post-proc│
       │ review  │ │ + PDF   │
       │  Opus   │ │ Sonnet  │
       └─────────┘ └────┬────┘
                        ▼
                   comic.pdf

This is what we do everywhere else. We take organizational knowledge, our understanding of AI, use our runtime to wrangle and make reliable multiple collaborating models, and ship the outcome, not the tooling.

We’re funded by friends - the wonderful folks at the General Partnership, Dhuna Ventures, and angels from Databricks and elsewhere, who believe that solving the problem of first-mile data ingest is where we can create more value than we capture. And a whole lot of it.

We split time between SF, NY and Singapore, but we call Singapore home. We’re always happy to trade hanks for conversations.