Agents don’t know how your company works.
Wrong. It only counted bonds Amazon issued.
Right. It found Amazon inside the CMBS and CLO collateral.
Sovara teaches your agents how your company works.
Build knowledge by correcting your agent.
Monitor.
Sovara records every run, so you can find the ones where the agent had to act on something it has never seen before.
Diagnose.
Understand what the agent did, whether it was wrong, and what it should have done instead.
Correct.
Teach the agent a lesson: a short note telling it what it should have done, so it gets it right the next time.
Sovara optimizes for human time.
Example: Sovara needs to know what your firm counts as exposure to a company.
cmbs_tenant_roll schema
Credit risk policy
Exposure methodology.docx
“Exposure is measured look-through to the underlying collateral.”
Lesson created
Risk agentOur Amazon exposure is $40M.
MariaThat is issuers only. You have to look through the CMBS tenant rolls.
Lesson created
#credit-risk
Lesson created
Sovara asks a person only when the answer exists nowhere else.
Case studies.
“For exposure questions in CMBS, consider look-through exposure and query the commercial buildings tenant table.”
Navigate siloed data and business processes.
We are collaborating with a financial services firm that invests across different credit products (corporate credit, CLOs, ABS, CMBS, RMBS, municipal bonds and related products). Historically each asset class has been run like a separate business inside the firm, with its own desk, workflows and data model. The desks do not talk to each other, which creates problems: the firm cannot easily see its true exposure to a company when that exposure spans several vehicles at once. Sovara learns each desk's internal processes and data models so their data can be stitched together.
FinanceBench
FinanceBench+
Accuracy boosts: questions over SEC filings.
Sovara can create agent instructions automatically from an existing labelled dataset. We took samples from the FinanceBench benchmark and let Sovara iterate on it to find an agent's common mistakes. We then created FinanceBench+, which combines and modifies FinanceBench examples. After learning from FinanceBench, Sovara boosts the agent's performance on FinanceBench+ from 80.6% to 96.8%.
Text-to-SQL translation for an enterprise database
Translating English questions to SQL.
We collaborated with a large organisation to enable Text-to-SQL translation for their database administrators. The database encodes years of tribal knowledge a general-purpose AI system cannot infer from schema alone. Similarly named tables such FCLT_HIST and FCLT_HIST_1represent different concepts, and domain terms like “levels” and “floors” are not interchangeable. Out-of-the-box tools such as Codex and Claude Code answered only ≤15% of queries correctly. After iterating with Sovara, their Text-to-SQL agent achieved 81.7%.
Try the Sovara desktop app for free.
Sovara CLI
1. Install the CLI
curl -fsSL https://apps.sovara-labs.com/cli/install.sh | sh2. Instrument your project
Run this command from your project directory:
sovara setupDesktop app
Other Linux builds: Debian / Ubuntu arm64 (.deb) · Fedora / Red Hat (.rpm) · Fedora / Red Hat arm64 (.rpm)