Know which AI does your firm's work to your standard before it touches a live deal. We build evals from your own tasks and documents, and score every model and agent you are considering.
Trusted by leading firms
Adjusted EBITDA bridge, buyer's view
6 checks, 300 points
Why evals
An eval answers them on your own work, with evidence.
Find the model, agent or vendor that does your work to your standard, at the lowest cost.
Measure how much your templates, playbooks and firm knowledge improve the output.
Catch regressions the moment a model, prompt or vendor changes.
A custom eval is a test set built from your firm's own work: real tasks, the documents behind them and checks written by your investors. Every AI system is graded on the models, memos and decks it delivers, against the standard your partners hold.
Case study
PE Bench
Private equity
Former private and growth equity investors turned five deals, from public take-privates to messy private data rooms, into graded tasks that run from diligence to the IC memo.
See the results →58
Graded tasks
5
Deals, public and private
80.5%
Compound's top score
Working together
Embed our team, grade your live work or build a private benchmark. Every path runs on the method behind PE Bench.
Embedded experts
Our former investors and research scientists work alongside your team to write tasks and checks for the work you are moving to AI.
Grounded in your deals
Ongoing evals on the work your team runs through AI, scored against checks your partners sign off on.
Built for selection
A fixed suite on your own deals. Run any model, agent or vendor on it and see exactly where each one breaks.
Cost and performance
Every system you test lands on one chart of score against cost for your work. Pick the cheapest one that clears your bar.
The system that clears your bar for the least.
Know what each task costs before you scale it.
Our team
Two of our founders were research scientists at Google Brain and DeepMind, where evaluating models was the job. Former private and growth equity investors write the checks.
Meet the team
Peter J. Liu
Co-founder
Research scientist, Google Brain and DeepMind
Yao Zhao
Co-founder
Research scientist, Google Brain and DeepMind
Yijia Liang
Co-founder
Former growth equity investor
Why Compound
The same method, graders and calibration we hold Compound to.
Every Excel model, Word memo and PowerPoint deck is recalculated in engines that match Microsoft Office and checked cell by cell.
Each task has a reference answer that must score 100% on its own checks before the suite grades anything.
Where prose needs judgement, the AI judge grades only after it agrees with your reviewers.
What you get
Built from your work. Yours to keep.
Tasks, documents, weighted checks and reference answers.
Score by task, consistency across runs, and cost and time per task.
Every missed check points to the step that failed.
Rerun on every new model, prompt or vendor and see what changed.
Coverage
Real briefs on real documents, graded on the file that comes back.
Questions across a full data room where the seller's documents disagree.
Flags FY25 revenue in the management deck that does not tie to the audited accounts.
A bridge from reported to adjusted EBITDA, built to your firm's add-back policy.
Rejects deal costs incurred before the board authorized the sale process.
Live Excel models on your template, recalculated and read cell by cell.
Drives forecast revenue from volume and price.
Investment committee memos and decks on your template, tied to the model.
Puts the price and the bridge to it on the first page.
Earnings reviews, estimate updates and filing comparisons for public equity teams.
Restates the prior year to the company's new segments.
Covenant tests, debt stacks, lender updates and LP reporting.
Computes leverage on the credit agreement's EBITDA definition.
Leaderboards test someone else's tasks against someone else's answers. Your firm needs a test of its own.
Leaderboards average thousands of questions your team never asks, and the top models sit within a few points of each other. Two models with the same headline score can differ widely on the task you care about, and on cost.
In BankerToolBench, built with 502 investment bankers, none of the nine models tested produced work the bankers would send to a client.
BankerToolBench, April 2026In the GAUGE study, one analyst's valuation model scored only 0.33 against another's as the answer key. Deal work needs checks that accept any defensible approach.
GAUGE, 2026Your add-back policy, model template, IC memo structure and the questions your partners always ask are what separate usable work from rework.
From first conversation to an eval suite you own, in four to six weeks.
A public benchmark ranks models on tasks someone else chose, against someone else's answers. A custom eval uses your tasks, documents and conventions, so its score predicts how an AI will perform on your next deal.
See PE Bench →Former research scientists from Google Brain and DeepMind, working with former private and growth equity investors: the team behind PE Bench. Your investors review every check before it scores anything.
Meet the team →Yes. Every system gets the same briefs, documents and time limit, is graded by the same checks, and its outputs sit side by side. On PE Bench, our own benchmark, Compound has the top overall score.
See the PE Bench results →No. The suite can run in your own AWS or GCP account, and Compound never trains on customer data. Your tasks and answer keys are never published or reused.
Security details →Checks test whether the call is defensible and supported by the documents. Where your firm has a policy, such as which add-backs it accepts, the check holds the agent to it.
How PE Bench grades judgement →Most checks read the delivered files directly. For prose, the AI judge is scored against your reviewers' verdicts on the same outputs, and grades nothing until it agrees with them.
Yes. Models are recalculated and read cell by cell, which catches pasted values and broken formulas. Decks are rendered and checked against your template and the model behind them.
How we rebuilt Office for agents →A first suite usually starts with a few dozen tasks from recent deals and known failures, enough to separate the systems you are considering. It grows as your team finds new failures worth testing.
Four to six weeks from first conversation to a reusable eval suite built on your firm's work.
Three ways: our expert team embeds with yours to write tasks and checks, we run live evals on the work your team sends through AI, or we build a private benchmark on your deals to compare models and vendors.
Yes. Tasks, checks, reference answers and every graded run are yours, and the suite reruns whenever a model, prompt or vendor changes.
The SEC's 2026 exam priorities say examiners will check that firms' statements about their use of AI are accurate. A dated record of what you tested, on which model and with what result, gives your compliance team evidence to point to.