The only AI your firm
can trust with its numbers
Frontier models and finance AI tools are built to sound right. Kepler is built to be right: code pulls and computes every number from your sources, and every answer checks out to the line.
Request DemoA deterministic engine
Code pulls and computes every number. The model never writes one.
A deep search index
Every document read in full and indexed before you ask.
Your firm's ontology
Your entities, definitions and relationships, resolved and sourced.
Line-level provenance
Every figure traces to the line it came from, with its formula.
Checked as it's written
Every citation verified against its source before you see it.
Any point in time
Every figure as filed or as restated, on any date.
How they
actually work
Ask these questions of any AI tool before you trust its numbers.
| Question | Frontier models on their own | Finance AI tools | |
|---|---|---|---|
| Where each number comes from | Written by the model | Written by the model | Pulled from the source and computed by code |
| Can the citation be wrong? | Yes: the model attaches it | Yes: the model attaches it | No: it's where the number was pulled from |
| What the citation points to | A document or a passage | A document or a passage | The exact line the fact came from |
| Finding the right facts in your documents | Only what fits in one prompt | Searches for passages when you ask, and can miss them | Every document read in full and indexed before you ask |
| When two sources disagree | Picks one, without telling you | Picks one, without telling you | Shows both, each with its source, and flags the conflict |
| Showing the calculation | Hidden inside the model | Hidden inside the model | Run as code, with every input shown |
| Asking the same question twice | Can give two different numbers | Can give two different numbers | The same number, every time |
| Using your firm's definitions | Re-typed into every prompt | Templates the model may not follow | Built into your ontology, applied every time |
| Checking before you see it | After the fact, if at all | After the fact, if at all | Every citation, while the answer is written |
| Figures as filed or as restated | Can't tell them apart | Can't tell them apart | Either, on any date |
Where other AI
goes wrong
What was Alphabet's operating margin in FY2025?
Alphabet's FY2025 operating margin was 31.4%.1
104-page PDF
× The model wrote 31.4%. Nothing in the filing says it.
What was Alphabet's operating margin in FY2025?
Alphabet's FY2025 operating margin was 32.0%.
129,039 ÷ 402,836 = 32.0%
Pulled from Item 8 by code, then computed. The same answer every time.
What was the company's total debt at quarter end?
Total debt was $10.9 billion.1
× The citation is real, but it's one line of two. Long-term debt is missing.
What was the company's total debt at quarter end?
Total debt was $35.3 billion.
10,883 + 24,410 = 35,293
Every component pulled and cited, then summed in code.
Does CoreWeave depend on a single customer?
No. CoreWeave serves a diversified base of AI labs and enterprises.1
Business overview
× Taken from the overview. The concentration is disclosed further in.
Does CoreWeave depend on a single customer?
Yes. Microsoft was about 67% of revenue in Q3 2025.
Found because every page was read before you asked.
Methods we publish
Questions
Why can't a frontier model be trusted with the numbers?
Because it writes them. A model predicts the most likely next digits, so a figure can be wrong and still look right, with a real source attached. Kepler never lets the model write a number: code pulls it from the source and does the math.
Don't finance AI tools already cite their sources?
They attach citations to what the model wrote, and a citation can prove a source exists without proving the number came from it. Data checked by people only covers public filings, never your own documents. Kepler pulls every number from the source with code, in your documents and in filings alike.
What makes Kepler different?
A deterministic engine for every number, a search index of every document read in full, your firm's ontology, line-level provenance, citations checked as the answer is written, and every figure as filed or as restated on any date.
Can we keep the AI tools we already use?
Yes. Kepler runs inside Claude as a verified connector, and any agent can use it through the Kepler MCP and API.