The IC memo you can defend — it audits sponsor models, sources every number, and lets arithmetic referee the bear case.
Memosa is an investment-memo platform I designed and built solo over roughly 20 months at EquityMultiple. It exists because IC memos were eating analyst weeks, and because I wanted to see whether the reliability problems everyone attributes to LLMs could be engineered around.
A commercial-real-estate investment committee approves deals worth tens of millions on the strength of one document: the IC memo. At EquityMultiple, producing one meant an analyst spending one to two weeks reconciling a sponsor's Excel workbook by hand — rent rolls, pro formas, debt schedules, sensitivity tabs — then hand-checking every number and writing the narrative. The model is the deal. A single circular reference or hardcoded override, unnoticed, can flip an investment decision.
A general-purpose chatbot is the wrong tool here. It hallucinates numbers, can't trace a claim back to the cell that produced it, and writes memos no committee can defend. The bar was never "draft faster." It was: every number sourced, every downside checked against the model's own math, and nothing going out below a defensible integrity grade. That means reading a spreadsheet as a program — a graph of formulas — not as text to summarize.
I reconstruct a sponsor's Excel model as a dependency graph (about 2.4M dependencies in a standard model, ~120k formulas across 13 tab families) and run 43 deterministic checks — circular references, hardcoded overrides, reserve gaps — before Memosa writes a word. Four ideas carry the design:
From EquityMultiple's deal flow:
Keeping this many moving parts coherent — the Excel dependency-graph parser, the deterministic referee, the entity graph, the multimodal retrieval, the real-time collaborative editor (operational transforms over Redis), the integrity and audit system, the investor-packet generator — came down to the same discipline the product enforces: the LLM drafts and explains; it never gets to be the thing you trust for a number.
Memosa is a proprietary EquityMultiple product, currently in early access. The architecture and metrics here are drawn from its public product documentation; the source is not open.