Case studies / AI
Finding lost revenue with private, on-premise AI
Print media company: printing, distribution, billing and collections, and advertising sales
Get the full case studyThe challenge
Revenue was being lost in the gaps between contracts, pricing, and billing: missed charges, incorrect prices, unbilled usage, and duplicate credits. The data that could prove it was not permitted to leave the premises.
The client's data center. Nothing leaves it.
- Analysts
- Flagged casesEvidence drill-downApprove or deny
- Controls
- Evidence gatingHuman approvalAudit logsRole-based access
- Detection
- Private LLMRetrieval (RAG)Rule checks
- Compute
- GPU inference serversModel gatewayVector store
- Data
- Account numbersProduct and pricingLine-item billingContracts
What we did
We built a private, on-premise LLM platform and connected it to account, product, and line-item billing data. Names, addresses, and payment details were kept outside the model. Retrieval across contracts and pricing documents works in concert with rule checks, and analysts review each flagged case alongside its evidence.
- NVIDIA
- Llama
- PyTorch
- Hugging Face
- LangChain
- Milvus
- Docker
- Kubernetes
- On-premise
- private LLM on GPU inference servers
- 3 data domains
- accounts, products, and billing line items
- Evidence-gated
- every finding traceable to source records
- Human in the loop
- before any money-impacting action
The result
Every finding is accompanied by the reason it was raised, the source records, a confidence level, and an estimated dollar impact. A person approves it before any money moves, so finance and audit can rely on it.