TokenBook

Inspiration

At CalHacks, most of us knew we wanted to work on a fintech problem, but hadn’t settled on one. A conversation with a Hyphenate engineer got us thinking: companies plan for hardware and raw-material costs by asking vendors what to expect. AI software is harder to plan for because the bill changes with how teams use different models. We wanted to explore how finance could get an earlier view of that

What it does

TokenBook forecasts AI spend by team and connects each estimate to the work behind it. It helps finance compare estimates with actuals, see budget impact, prepare an accrual for controller approval, and compare model options by workload and cost.

How we built it

We built TokenBook with React, TypeScript, and Node.js, bringing Slack, GitHub, Jira, and Gmail context into a finance workflow. Anthropic’s Claude SDK helps interpret work signals; structured checks validate candidate changes, while code handles forecast and accrual calculations. Apify collects model pricing for cost comparisons. We refined the extraction prompts and tested them against cases involving conflicting dates, duplicate claims, retractions, and bot messages.

Challenges we ran into

Teams and individuals use different models for different kinds of work. We had to represent those workloads, understand what teams are planning, and connect that context to a useful cost estimate without treating every mention as extra usage.

Accomplishments that we're proud of

Most of us came to CalHacks for our first hackathon with an interest in fintech and being able to trace purchases, but no defined idea. We’re proud that a conversation with a Hyphenate engineer helped us spot a new planning problem in AI spend and turn it into a working product direction.

What we learned

A forecast needs to be explainable as well as accurate. Finance should be able to trace an estimate to its source, understand the assumptions, and compare it with actual spend. Model recommendations also need to account for what the team is doing, not just which model is cheapest.

What's next for TokenBook

We’d connect more company context, including cloud usage and billing, headcount, business operations, and engineering workflows. Then we’d keep validating estimates against actual spend and improve workload-specific model recommendations.