01 / Strategy
Cut surface area, not coherence
Broad intelligence product, reduced to three two-week sprints.
PC Deals, comparison, Portfolio Benchmark, and Shared Requests.
PRODUCT · AI · FINANCIAL INTELLIGENCE
NEW YORK, NY ↗
I build products for complicated problems that do not fit neatly into a prompt.
I’m Carlos Abreu, a product manager working at the intersection of AI, financial data, legal analysis, and enterprise software. Most of my work begins with an ugly problem: hundreds of fields, domain rules nobody has written down, legacy systems, expert judgment, incomplete data, and a deadline. I turn that ambiguity into a product.
The product is the whole system
01—06- User goal
- Submit the correct agreement.
- System responsibility
- Preserve source identity and permissions.
- Primary risk
- Duplicate, missing, or non-entitled source.
- Product principle
- Deterministic identity comes before similarity.
- End-to-end platform ownership
- Human-in-the-loop AI
- Private credit & covenants
- Structured data & APIs
- Evidence & evaluation
- QA, UAT & release
01 / Product judgment
The decisions are the work.
02 / Data model
One deal, one current view, history underneath
One deal can contain term sheets, preliminary reports, final agreements, and amendments.
Use the latest published report and representative benchmark tranche.
03 / Data model
Do not aggregate tranches when semantics are undefined
Tranches have different covenant semantics.
Show the representative benchmark tranche and its actual name.
04 / Data model
Separate identity from occurrence
Report-level TrancheIDs change even when the economic facility persists.
DealID → TrancheLineageID → report-level TrancheID occurrences.
05 / AI & trust
Automate candidate discovery before automating judgment
Borrower name alone cannot determine whether an agreement belongs to an existing deal.
Automate candidate discovery before judgment.
06 / AI & trust
Make human review become exception handling over time
Default human review limits scale, but unsupported autonomy creates risk.
Capture corrections, measure errors, calibrate, then automate high-confidence cases.
07 / AI & trust
Private data boundaries belong in the architecture
Private agreements are client-owned; global AI search can cross boundaries.
Client-scoped sources and anonymized market aggregates.
08 / Data model
Missing is not No
Missing values can silently become negative observations.
10 Yes / (10 Yes + 5 No) = 66.7%; exclude 5 Missing.
09 / Workflow
Put the hard gate at the consequential action
Dependent required fields initially blocked saving partial work.
Put the hard gate at the consequential action.
10 / Workflow
Complete for what?
A generic Complete flag collapses multiple readiness states.
Extraction complete; scoring 100%; report 87%; publication blocked.
02 / Selected work
Problems I’ve been trusted to solve
01 / Expert workflowShipped
Covenant Studio
How do you turn a 200+ page agreement into trustworthy structured intelligence?
Inspect the case study02 / Data productIn development
Covenant Intelligence
How do you turn lawyer-reviewed reports into a client-facing data product?
Inspect the case study03 / Data productProduct concept
Deal and Tranche Identity
How do you know two records are the same economic deal or tranche?
Inspect the case study04 / AI & trustProduct concept
Agentic Private Credit
How do you let AI act on private financial data without crossing the client boundary?
Inspect the case study03 / The details matter
I like products where the edge cases are the product.
Is zero actually zero?
Blank, zero, No Provision, Not Specified, and Not Extracted can mean five different things.
Where does the user go after an error?
Success means reaching the result, not receiving a success message.
What happens to my work?
An expert should not re-enter a section because one validation failed.
Which version am I looking at?
Preliminary, final, and amended agreements should not silently collapse into the deal.
Can this user see this?
Test navigation, direct URLs, search, documents, reports, APIs, benchmarks, and AI retrieval.
What if the happy path is false?
Duplicate documents, closed tranches, unresolved matches, partial extraction, and missing evidence.
What does the denominator mean?
Five missing agreements counted as No materially change the analytical claim.
Let’s connect