Faster extraction is not the same as a structured decision.
Endex AI does one thing very well: it parses private company financials and accelerates the modeling work that comes after. Capital Refinery uses extraction as the input layer to a much larger problem — keeping the IC decision alive across the life of the position. Different scope, different goal, complementary tools.
What Endex does well, and where the line is.
What Endex is built for
AI-driven financial extraction and modeling on private company documents.
Pulling structured financials out of CIMs, financial statements, and management decks. Generating draft models and analyst-grade synthesis on the contents. For deal teams under time pressure during diligence, it compresses the time from raw documents to a working model — and that compression is the entire point of the product.
What it is not built for
Holding the IC decision as a structured testable record after close.
Endex is an extraction-and-modeling accelerator. The output is a model and a set of structured numbers. It is not a system of record for the decision the team made on the basis of those numbers. The thesis, the conditions, the operator data flowing in three years later — none of that is in the Endex data model.
Where Capital Refinery sits next to it
Underneath the IC stage, on the other side of close.
Capital Refinery's extraction kernel is deterministic-first with a thin LLM adjudication layer — a different architectural choice than Endex's generative-first approach. But the bigger difference is downstream: Capital Refinery binds the extracted figures to an IC decision and re-tests them continuously. Endex hands off the model and walks away.
Where the lines fall.
| Capability | Endex AI | Capital Refinery |
|---|---|---|
| Financial extraction from CIMs and statements | Strong — core capability | Deterministic-first kernel with LLM adjudication |
| Draft model generation | Strong | Yes — but as input to the structured record |
| Structured investment record at IC | No — output is a model file | Yes — bound at approval with thesis and conditions |
| Operator data → entry assumption mapping | No | Yes — the data model |
| Decision validity scoring | No | Yes — live across the position |
| Provenance trail to source page | Citation-style | Every figure has a candidate ID and method |
| Covenant + breach probability for credit | No | Yes — for credit positions |
If your team already runs Endex for extraction, that is fine.
The natural next question is what governs the decision after the model is built. Bring us a deal you closed last quarter.