The deliverable isn't just the add-on — it's the judgment that produced it: a defensible point of view about where a solo builder can win in institutional middle-office tooling.

What was delivered

A working, bound Google Sheets add-on that reconciles an internal book against a broker file, classifies each exception by root cause with a deterministic engine, surfaces one-sided population breaks as first-class, and uses Gemini to draft the counterparty dispute email from that classified output — the full MVP cut-line, running on synthetic OTC variation-margin data.

More importantly, the project delivered a defensible point of view: that the opening in institutional middle-office tooling isn't a new capability (they're all owned) but the last-mile human loop, and that the way to occupy it is a thin, auditable layer inside the tool the desk already uses — with the AI writing the dispute email and a deterministic engine owning the classification.

What it demonstrates

Framed for what this case study is actually for — evidence of product judgment, not a shipped company:

• Competitive and market literacy. A full-lifecycle teardown that names the real incumbents and kills weak ideas on evidence, plus the buyer economics that explain why small desks are underserved. Most portfolios have no discovery of this depth.
• Architecture judgment. The rules-vs-LLM split — the AI writes the email, the engine decides what counts as a break — is what makes an AI tool trustworthy in a workflow where a wrong number costs money. Deciding which job the model does is the senior call.
• Domain fluency. Population breaks, the CSA Minimum Transfer Amount, pay-undisputed-hold-disputed: specifics a generalist can't fake.
• Build capability. An actual functioning add-on, from a first-time build, including the AI integration.

Honest limitations

Stated plainly, because owning them is part of the signal:

• Synthetic data only; no live integration. Usability testing was real but modest — 23 people ran the tool, 19 completed the core flow (reconcile → draft), and 6 finished the full 6-question survey; directional and qualitative, not statistically significant.
• Exact-match headers and tab names; no currency normalization; tolerance fixed in code rather than an interactive control.
• Built on Google Sheets, where the engine runs server-side; the target Excel build would keep matching local, with only minimal break summaries reaching a zero-retention LLM endpoint.
• Vendor and market facts reflect research as of mid-2026 and warrant re-verification before public use.

What's next