A small team, running like a much larger fintech.
A lending marketplace with five product surfaces, five stakeholder groups, and the operational load of a company several times its size. We embedded Otto, our AI operations and engineering layer, inside it.
The load of a large fintech. The headcount of a small one.
The business runs a consumer lending marketplace: consumer application forms, a matching engine, partner and referral flows, platform and admin tools, and the backend services beneath them. Five surfaces, one small team.
Around those surfaces sit lenders, partners, internal operations, engineering, and leadership, each with their own reporting, QA, deployment, and incident demands. Partner-specific requests arrive weekly. Funnels need analysis. Production and staging need coordination.
Speed matters. But mistakes in financial data operations are expensive. The team could not buy speed by accepting risk. It needed both.
Otto, inside the system.
Otto is the company’s embedded AI operations and engineering assistant. It connects to the repositories, production and staging systems, reporting workflows, team channels, scheduled jobs, and the internal knowledge base, and works across all of them to analyse, build, test, monitor, and report.
Otto does not run unsupervised. It operates with human oversight, staging-first deployment rules, and clear approval gates for sensitive actions. That is the design, not a limitation.
Five duties. One operator.
Reviews code and pull requests. Investigates bugs. Coordinates staging-first deploys. Checks branches, diffs, CI, and deployment state. Produces implementation plans before larger changes.
Pulls funnel metrics. Builds end-of-month and month-to-date reports. Produces partner-specific performance views: applications, matched leads, conversions, referral clicks, revenue. Flags anomalies and data gaps.
Runs a daily engineering pulse. Monitors health and deployment status. Coordinates QA requests. Summarises what changed and what needs review.
Maintains a persistent company wiki: decisions, architecture notes, partner context, reporting assumptions, operational runbooks. The team stops re-explaining itself.
Low-risk display and reporting changes move fast. Database migrations, production pushes, lender configuration, and sensitive data operations require review and approval. Staging is the default proving ground.
What changed.
Before
- Reporting meant manual SQL, exports, and reformatting.
- Context lived in people’s heads and scattered chats.
- Engineering issues needed repeated handover.
- QA and staging checks were ad hoc.
- Partner requests created repeated one-off work.
- Leadership chased status updates manually.
After
- Repeatable reports generate from standard workflows.
- Operational context is retained in the wiki.
- Engineers receive complete tickets with context attached.
- Staging and deployment conventions are enforced.
- Partner-specific analysis is produced on demand.
- Leadership asks operational questions in natural language.
Five entries from the record.
An internal end-of-month report existed. Otto converted it into individual partner-facing reports: applications, matched and completed, conversions, lender clicks, referral value, and a short insights section per partner.
Every report is reviewed by a person before it leaves the building.
Otto holds the company’s branch and deployment conventions: staging deploys first, production requires explicit approval, every time. The conventions are enforced by the operator that does the work.
This turns AI from a risky coder into a controlled operator.
Each morning Otto summarises the engineering board, staged QA, overnight development, deployment health, and the backlog. Leadership reads the state of the operation without chasing a single engineer.
QA requests are structured and posted to the QA channel. Bugs and issues route back into the engineering workflow with context attached, which matters on the consumer forms and platform surfaces, where partner-facing detail is the product.
Otto maintains company-specific knowledge: architecture notes, partner identifiers, deployment conventions, report definitions, project plans. Switching projects no longer means losing the thread.
The numbers will be the run log’s, not ours.
Outcome figures for report turnaround, investigation time, QA coverage, and anomaly detection are being measured against the pre-Otto baseline now.
We publish verified numbers or none. Figures from this engagement are available on request once the measurement period closes.