AI
How I build with AI
Four rules, and every Hemrock project follows them.
Rule one
AI for reasoning
Use the model for what it is good at: understanding context, reading messy documents, planning, explaining, spotting issues, choosing a tool, and proposing an action.
- Understanding context
- Reading messy documents
- Planning
- Explaining
- Spotting issues
- Choosing a tool
- Proposing an action
On this site it is the Ask AI sidebar, which reads the docs and decides which engine to call. In Portfolio it is the intake that turns a company update into metrics for review.
Rule two
Deterministic software for math
Anything where the same inputs must produce the same outputs runs in code, not in a prompt.
A model doing arithmetic in its head will eventually miscount an option pool or convert a SAFE on the wrong basis, and it won’t tell you. So the cap table, exit waterfall, and fund economics engines do that work, and the AI decides what to ask them.
The engines are tested against the spreadsheet models they came from, and the in-browser tools, the API, and the MCP server all run the same core. A cap table an agent computes matches the one in your workbook.
Context, for the judgment
Primers, task prompts, sanity checks, and concept guides that teach your AI how a model is structured and how a calculation works.
Compute, for the math
The engines themselves, callable directly. Handing the math over removes the failure mode instead of hoping a prompt catches it.
Rule three
State for continuity
Finance is a running record, not a chat. Transactions, accounts, assumptions, policies, entities, decisions, and history have to persist and be inspectable.
- Transactions
- Accounts
- Assumptions
- Policies
- Entities
- Decisions
- History
Portfolio is built this way: every metric keeps the document it came from, a person confirms what it means, and the LP capital accounts are derived from a real ledger.
Rule four
Humans for accountability
AI can prepare, analyze, recommend, and automate. A person approves. In Portfolio every extracted metric waits in a review queue, and nothing reaches a report until someone confirms it.
The models are good. Someone still has to be answerable for the number.
In practice
Where this shows up
Extraction with review, a real ledger, and LP reporting you can trace.
Cap tables, waterfalls, and fund economics, exact every time.
I help teams decide where AI belongs in their stack, then build it.
Build with these tools live, and keep the math.
Most people use AI to build a financial model from a blank sheet. I think that’s backwards.
Start from something proven, let the AI do the reading and the line-level edits, and call an engine for any number that has to be right. That is how the models on this site are built and how Portfolio runs.