Replacing a 2013-Era Collections Suite with a Multi-Tenant Operations Platform
A debt-recovery business had run every workflow inside a legacy collections product since 2013. We rebuilt the operation around its own process: APIs instead of file drops, and plain-language reporting on an on-premise LLM.
- Service
- Business Transformation
- Industry
- Revenue Cycle & Finance
- Region
- United States
- Engagement
- Lab-type ODC (dedicated team)
Outcomes
- Vendor lock-in removed
- File drops replaced by APIs
- Plain-language reporting on-prem
Stack
Python · PostgreSQL · On-prem LLM · REST APIs · Containers
Written byOriental Nippon Team
A mid-sized US recovery business works on behalf of healthcare providers. Its whole operation, from data intake and account handling to agent work and client reporting, ran on a legacy collections suite whose workflows had been fixed since 2013.
The software worked in the narrow sense that it did not fall over. But the business could not change how it worked without going back to the vendor, and it had stopped trying.
The problem
- Frozen workflows. Adaptations had not been revised in over a decade.
- Vendor dependency. Any process change meant a vendor request.
- File-based data movement. No APIs, only rigid file drops in and out.
- Static reporting. Anything outside the shipped reports went to a central team's queue.
- No room for hosted AI. Client-sensitive data ruled it out.
| Before | After | |
|---|---|---|
| Workflow changes | Vendor request | Configurable by the business |
| Data in and out | File drops | REST APIs |
| Client separation | By convention | Enforced at platform level |
| Reporting | Static set plus a queue | Ask in plain language, on-premise LLM |
What we built
A tailored operations platform
Account handling, work allocation, client and vendor interfaces, and reporting.
Multi-tenancy
Each client's data separated at the platform level, not by convention.
API-based data flows
Replacing the file-drop model.
On-premise AI reporting
A language model deployed on the client's server, with voice support, so reports can be asked for in plain language and no data leaves the network.
Technology and the role it played
| Technology | Role in this project |
|---|---|
| Python | Application logic for account handling and allocation |
| PostgreSQL | Tenant-separated operational data |
| REST APIs | Data exchange with clients instead of file drops |
| On-premise LLM | Plain-language and voice reporting inside the network |
| Containers | Repeatable deployment across environments |
How we delivered it
Process came before software. We mapped the work with the people doing it and separated real requirements from habits the old system had forced on them.
- Process mapping. Current state, fit-gap, target process and tenancy design.
- Foundation. Environments, security model, tenancy structure and deployment pipelines.
- Area by area. Each functional area demonstrated to real users and signed off before the next began.
- Parallel operation. The legacy suite stayed available while each area moved.
- Stabilise and hand over. Documentation, role-based guides and a backlog of deliberately deferred items.
01Discovery and process mapping
3–4 wks
Current state, fit-gap, target process, tenancy design
02Foundation
3–4 wks
Environments, security, multi-tenancy, pipelines
03Core build waves
12–20 wks
Account handling, allocation, client interfaces, reporting
04On-premise AI layer
4–6 wks
Local model deployment, schema grounding, voice interface
05Cutover and stabilisation
4 wks
Parallel run, sign-off, documentation, hand-over
Risks and how we handled them
| Risk | How we handled it |
|---|---|
| Rebuilding old habits in new software | Process mapped with staff first, separating needs from legacy habits |
| A big-bang cutover | Area-by-area sign-off with the legacy suite still available |
| Client data mixing | Multi-tenancy enforced at the platform level |
| Scope creep | A documented backlog of deliberately deferred items |
Outcome
The business can change its own processes again. Workflows that had been static since 2013 are configurable, data moves over APIs instead of files, and reports can be requested in plain language instead of from a queue.
Whether that turns into a measurable commercial gain is something the client will know better than we will. We would rather wait for their number than invent one.
What the client keeps
- The platform and its source
- The tenancy and security model, documented
- API specifications
- Process maps and the discovery decision log
- Role-based user guides
Sources and further reading
- Strangler fig application, Martin Fowler
- Multi-Tenant Security Cheat Sheet, OWASP
- Row security policies, PostgreSQL documentation
Facing something similar? This is the kind of work our Enterprise & Digital Transformation team delivers every day. Talk to us about your stack.
