Portfolio
Different domains, same motion. Start with something that matters. Organize only what it needs. Put it into use. Keep what you learned.
The details change. The discipline does not: make what is known clear, keep the unknowns visible, and leave a better starting point behind.
In progress
These are active examples of the same motion. Running means there is something usable today. Building means we are still putting the pieces together.
01 / Running
2026-08-13
Procurement
documents → narrow extraction → deterministic score → quoted evidence → portfolio re-score
A working contract-risk system separates reading from judgment. Change a policy lever and the portfolio re-scores without another model call.
02 / Building
2026-08-12
Banking / community development
five public data families → entity resolution → evaluation corpus → peer evidence → geography-level signal
The source universe is mapped. The hard part is joining incompatible identifiers, historical geography, lending data, and regulator narratives without losing the as-of date.
03 / Running
2026-08
Insurance / structured finance
statutory filings → layout fingerprint → deterministic parse → reconciliation gates → security and property holdings
The parser accepts a value only when the filing evidence and reconciliation checks support it. Partial and refused parses remain visible instead of being guessed through.
04 / Building
2026-07
Structured credit
fund filings + daily holdings + statutory sources → identity graph → holder observations → change signal
The visible holder base is scattered across incompatible sources. The work is to reconstruct the observable slice and make coverage explicit.
The motion
You do not need to organize everything first. Each piece starts with one useful question and only the information needed to move it forward.
For the practitioner
Under the hood
Every problem should leave the practice more capable than it found it.
Questions
Some became products. Some became recurring systems. Some simply made the boundary between what was known and what was not much clearer.
Delivered examples
The reader should not need to know the domain first. Each example starts with the practical need, then shows what we put into use and what became easier to see or do afterward.
Financial data
People can get an answer quickly. The harder part is knowing where the number came from and whether it still holds. We built a question-answering service over public financial data that keeps the source and date with the answer. What stayed useful was the evidence, not just the chat.
Structured finance
The information needed to understand a deal is public, but it is spread across filings and reports. We built DealCharts to organize that information into searchable deal, fund, and BDC pages with the source trail attached. Once the information is organized, people and models have a much better place to start.
Public SEC filings and remittance reports
Deals, funds, BDCs, and downloadable datasets
Procurement
A buyer needs more than a vendor explanation before accepting an increase. We built a signal from public energy, producer-price, and industry-use data to show when a resin increase may hold. The useful lesson was that public data can provide a starting position quickly, while the team's own contract data can sharpen it later.
Henry Hub, BLS PPI, and WTI, all public
213 scored events, 56.8% directional match
Contract review
Teams repeatedly review the same contracts against changing policies. We built a system that reads executed statements of work, scores each one pass, review, or fail, and keeps the quoted evidence with the decision. Change a policy threshold and the whole set re-scores without asking a model to reinterpret everything.
Deterministic verdict
review · 0.55
quote evidence
source unchanged
policy re-score without another model call
Ownership
The question is simple. The public record is not. We traced one class of a live auto ABS deal across prospectus, servicing, fund-holdings, and risk-retention filings. We can name some ownership and show exactly where public visibility stops. The important part is keeping what is known and what is still unknown separate.
SEC 424B5, 10-D, N-PORT, and Reg RR
One class traced, sources still being added
SEC filings
Teams were asking AI questions that depended on current SEC filing context. We built a bounded EDGAR tool surface so compatible agents can query filing metadata directly. The useful shift is simple: the model can reach the source material when it needs it instead of starting from another manual upload.
SEC EDGAR filing metadata
MCP-compatible clients
Verification
A claim from a model, memo, or article is easy to repeat and harder to verify. We built a public verification service that checks supported CMBS claims against SEC filings, links to the sources, and says when it cannot verify something. Verification becomes a repeatable step instead of another research project.
Public verification surface
https://receipts.cmdrvl.com/mcpclaude mcp add --transport http receipts https://receipts.cmdrvl.com/mcpcodex mcp add receipts --url https://receipts.cmdrvl.com/mcpConnect, call quickstart, then send the text to verify_claims. Call coverage before reading an UNVERIFIED result.
Point in time
Historical financial questions go wrong when today's knowledge leaks into yesterday's answer. We built a benchmark that asks what value, citation, and data date were actually available at the time, and expects the system to decline when the fact was not yet knowable. The unknown is part of the answer.
Point-in-time check
What was knowable on that date?
decline: post-cessation
96-task demo fixture
Tell us what it is and what you already have. We will find the smallest useful place to start, and you keep what gets built.
Book a call