The platform behind every Kepler answer, from the first read of a filing to the cited number on your screen.
Read the researchThe model writes the sentence. Code produces every number in it.
“What was Alphabet's operating margin in FY2025?”
3 figures, 0123 cited, 0 written by the model
Solid bars ran as code. The hatched bar is the model, writing the sentence. Every step runs inside your firm's isolated environment and is logged.
Sources: Stop Paying to Forget (Read, Retrieve) and Anthropic's Kepler case study (Route).
From the source document to the screen, each layer is built so an answer can be checked. Govern wraps all ten.
Public filings and market data arrive already loaded. Your firm's systems connect read-only and stay in sync, and every document keeps its provenance from the first byte.
Extraction models we train read each document the moment it lands: every fact with its value, scale and unit, every entity, table and relationship, each kept beside the passage it came from.
Companies, funds, people and filers resolve into one graph. Every edge is a sourced claim with its verbatim sentence, and the graph keeps time: what was true as filed, as restated, or on any date.
Hybrid retrieval over each firm's own index, scoped by entity, period and permission. A question about something already read is a lookup, not another read.
Statements assemble from the filings' own XBRL with no model in the path. Every calculation runs as code over cited inputs, and restatement histories are solved as an integer program.
The model names facts by reference and the platform renders the filed value. Every citation is checked against its source while the answer streams, and one that doesn't match is stripped before you see it.
Every request starts with a plan. Agents run as typed state machines that unlock only the tools each step needs, specialists work in parallel, and every thread survives a restart.
Tiered routing across frontier, open-weight and firm-hosted models. Small models we train do the reading, a frontier model handles the language, and any of them can be swapped: the evals decide what ships.
Your definitions compile into skills that are guarded, reviewed and attributed. Every correction becomes a label, models retrain privately for your firm, and double-entry evals gate every change.
The Kepler app, Excel, PowerPoint and Word, Claude and any MCP client, and our API for your own agents and systems. Every surface carries the same citations.
Each firm's data is isolated by identity policy, not just application code. Agents execute on a private network with no public ingress, every event lands in a sequenced log with full lineage, and we deploy in our cloud or yours.
The same work the trace shows, one layer at a time: sources syncing, a filing read in full, a company resolved and a model swapped.
Public data arrives loaded; your systems connect and sync.
Each fact is kept beside the passage it came from, so the next question is a lookup, not another read.
Each fact lands beside the passage it came from.
Four spellings, four sources, one record with your fields.
A better model swaps in. Workflows and numbers hold.
Every relationship we read is one filer's claim, kept with the sentence that states it.
Explore the ontology“NVIDIA is an investor in CoreWeave and supplies its GPUs.”
Statement histories rebuilt from the periods filings repeat, solved as an integer program with no language model.
ReadResearch noteRead every document once, keep every number, name and relationship, and answer the next question with ordinary software.
ReadResearch noteWhy better models make mistakes harder to see, and why we give them less to do with every upgrade.
Read