# Explore the scenario. Protect the record.

Ask “what gets this fund to 2.0x?” in plain language. 1,000+ scenarios run, none of them able to touch the ledger.

Company: VectorShift

Contribution: Engineering & product

Company tenure: VectorShift · Jul 2025–present. This is not a claim about individual project dates.

Context: Scenario changes never mutate the ledger

## TL;DR

Problem: To ask “what gets this fund to 2.0x?”, an investor had to rebuild the model in Excel.

My call: Let them ask in plain language, and keep every scenario architecturally separate from the ledger.

Outcome: 1,000+ scenarios run without touching the underlying ledger.

- Scenarios run: 1,000+
- Scenario writes to the ledger: 0
- How a scenario starts now: Excel → a question

## Context

- Company: VectorShift (YC S23)
- My role: Built the scenario layer on top of Portfolio Engine
- Users: Investors planning exits and return targets
- Depends on: Portfolio Engine as the source of truth

Conceptual separation boundary. The scenario layer reads the record, but scenario changes do not write back to the ledger.

## The problem

Exit planning is a chain of what-ifs. Sell this company a year earlier. Assume a lower multiple. Hold that one longer.

Each of those questions meant rebuilding the fund model in Excel.

That made scenario work slow, owned by whoever built the spreadsheet, and hard to check against the real portfolio numbers once the data had been exported.

Evidence: Investors were already asking return-target questions like “what gets this fund to 2.0x?” Exported models drifted from the system of record as soon as they left it.

## Constraints & non-goals

Scenarios must never mutate the underlying ledger.

The scenario layer starts from Portfolio Engine, so every scenario begins from sourced numbers.

Every answer has to make clear what is hypothetical, so a scenario is never mistaken for a reported figure.

Out of scope: Replacing the fund’s formal valuation process. Writing scenario results back into the record. Forecasting market multiples on the investor’s behalf.

## Options I weighed

Rejected: Export to a spreadsheet template Faster than rebuilding, but it copies the data out, and the drift problem comes straight back.

Rejected: Let the assistant edit assumptions in the live record One misread sentence could change ground truth. No convenience is worth that in a financial record.

Chosen: A separate scenario layer that reads the ledger and holds its own assumptions Investors can explore freely, because nothing they do in a scenario can corrupt the record.

## The spec

Users: Partners and principals planning exits. Fund finance teams supporting return analysis.

Requirements: Investors ask scenario questions in natural language. Scenarios read current figures from Portfolio Engine. Scenario changes live only in the scenario layer. Each answer lists the assumptions that differ from the record.

Acceptance criteria: No scenario operation can write to the ledger. Every result shows changed assumptions next to the baseline. Scenarios can be saved, revisited and compared.

Guardrails: Ledger integrity: zero scenario writes to ground truth. Scenario results are always labelled as hypothetical.

## What shipped

1. Ask. An investor asks a question like “what gets this fund to 2.0x?”

2. Read the baseline. The scenario layer reads sourced figures from Portfolio Engine.

3. Model. Assumptions change inside the scenario layer only.

4. Compare. The result is shown against the baseline. The ledger is unchanged.

## What changed

More than 1,000 scenarios have been run in natural language.

None of them touched the underlying ledger.

Scenario work stopped depending on whoever owned the spreadsheet.

Scenario count shows usage. I don’t have a measured time saving per scenario to report.

## Reflection

For AI products, the boundary is part of the experience. People explore with confidence only when they know what an interaction can change.

If I did it again, I would make the hypothetical state even more visible from day one. Clear labelling is cheap, and it stops a screenshot of a scenario being read as a reported number.

The takeaway: Separate hypothetical changes from financial ground truth.

Last updated: 2026-09-13