Latest Insights
Real outputs from Graam's agent pipeline. Click any report to read the full analysis.
Graam Is Your Credit Analyst
What it reads, what it already knows, what it produces, and three things it did this month
Graam reads your deal documents, loan tapes and data rooms, joins them with public and agency data, and does the work a desk hands to an analyst: the IC memo, the diligence on a tape, the structure sized to rating criteria, the performance report. Here is what it is, with three runs from this month.
Commoditizing Cashflow Models, Starting With EDGAR Filings
In 3 minutes, Graam's agents read SEC filings and build a cashflow model on their own, validated against published sensitivity tables and trustee reports, so the model checks the trustee as much as the trustee checks the model. Download the model and run it on graam-flows, the open-source cashflow engine.
An agent can now do the work of a cashflow modeler. Type the name of a public deal and Graam reads the prospectus, builds the waterfall, ties it to the prospectus's WAL tables, and replays it against every trustee statement since closing. WFCM 2018-C44: 754 pages read, 19 classes, 45 of 45 WAL cells, 100 trustee statements, zero unexplained breaks, all on an open-source engine.
A CMBS Deal, Read From EDGAR
GSMS 2017-GS8: the prospectus, the loans, nine years of remittances, and a WAL table that ties
Give Graam the name of a public CMBS deal and it finds the filings on EDGAR, reads the 731-page prospectus, rebuilds the structure, ties the model to the prospectus's own WAL tables, and reads nine years of monthly loan-level remittances to say where the deal stands today: 35 loans, $901 million, 14.8% in special servicing, and a 2027 maturity wall.
Diligence a Loan Tape Against County Records
302 synthetic loans on real parcels, one question, and a scored result
A loan tape with planted problems was built on real Texas and Arizona parcels and handed to Graam with one instruction: run due diligence. Every property was matched to its county record, appraisals were checked against the county market value, owner-occupied loans were checked against the owner of record, and each flagged loan was judged on its own facts. The post walks the property section, grades the run against the answer key, and shows the rest of the report.
A Neural Mind. A Symbolic Body.
Why structured credit needs neurosymbolic AI
Structured credit has always made you choose: tools flexible enough to answer any question (the spreadsheet) or rigorous enough to trust (the vendor engine) — never both. Language models promised flexibility but fabricate. Graam is neurosymbolic: a neural code writer that can only act through a symbolic engine and type system. The neural half decides and writes the code; the symbolic half does the math and refuses to let the model author a number. It's the one architecture that delivers accuracy, reproducibility, and auditability without giving up the flexibility of plain language.
What Stops an AI From Making Up a Number
Provenance, tiers, and a forgery guard
A deterministic engine fixes the math — it does nothing about the reporting. A model with a perfect calculator can still write an unsourced number into a term sheet. In Graam every figure is a typed object carrying frozen provenance, tiered by how strongly its origin is guaranteed: deterministic figures can only be minted inside the engine, and the model may not author them. Derived and stated values are never blurred, and when a value can't be sourced the system refuses rather than reaching for a plausible default.
Structuring a Securitization from a Loan Tape — Reproducibly
From a $600mm consumer tape to a priced, stress-tested capital stack
Starting from a $600mm consumer-loan tape, the agent stratified the pool by tier, sized an 8-tranche capital stack to a rating-agency credit-enhancement ladder with exact coupons, priced the bonds through the waterfall, and stressed cumulative losses past the 15.72% worst case — the subordinate tranches take real principal writedowns while the senior stack stays whole. Every figure traces to a deterministic primitive call or the tape, and the whole deal reruns from an open-source engine.
Open Sourcing the Execution Layer of Structured Credit
From CDI files to executable deals
We're open sourcing graam-flows — a deterministic cashflow engine that replaces proprietary CDI files with human-readable, executable deal representations. Collateral projections, waterfall execution, and bond analytics — no vendor lock-in required.
IC Report on NQM Securitization
From prospectus upload to full IC package in minutes
I generated an IC-style report from an offering doc in minutes — the agent decomposed a 285-page prospectus, built a 10-tranche cashflow waterfall, validated it against the flows engine, extracted and stratified collateral, queried GSE performance data to build S-curves and loss forecasts, ran scenarios with NQM stress multipliers, and produced tranche-level yields, WALs, and breakeven analysis.
IC Report on Subprime Auto ABS
From prospectus upload to full IC package in minutes
The agent parsed a subprime auto ABS prospectus, built the waterfall, stratified the loan tape by FICO, APR, and term, queried EDGAR shelf performance with vintage-over-vintage comparisons, constructed loss curves from empirical data, ran base, stress, and severe scenarios, and produced tranche-level yields, WALs, and breakeven CDR analysis.
Structuring a Subprime Auto Loan Sale
From collateral tape to rated capital stack and preliminary prospectus
Starting from pool characteristics, the agent built a 6-tranche sequential-pay capital stack with a turbo mechanism, sized credit enhancement against loss assumptions, validated the waterfall through the flows engine across 58 cashflow periods, and generated a preliminary offering supplement — cover page through risk factors.