Provider abstraction and switching
One interface for the model layer, so AWS Bedrock, OpenAI, Anthropic or a hosted open model can be swapped by configuration. We moved Notermed entirely from OpenAI to Bedrock this way.
Most companies don’t need a new AI product. They need the system they already have to summarize, classify, draft, search or extract. CS2’s AI integration services add LLM features to an existing product or internal system for businesses in Canada and the US, behind a provider abstraction, with privacy boundaries, cost and latency budgets, and a rollout you can switch off in a minute.
Summaries, smart search, drafted replies. The feature list is clear. The question is how to add it without destabilizing the thing that already sells.
Invoices, claims, applications, support tickets. Someone reads each one and types it into a system. An LLM can do the first pass, if the output is structured and checked.
The provider changed prices, or compliance now requires a different one. You need an abstraction layer and a safe way to switch.
The feature worked, but it took too long or cost more than it saved. Running it for real needs a budget for both, set per feature.
Health, financial or personal data can only reach providers covered by the right agreement. The integration has to enforce that in code, not only in a policy document.
One interface for the model layer, so AWS Bedrock, OpenAI, Anthropic or a hosted open model can be swapped by configuration. We moved Notermed entirely from OpenAI to Bedrock this way.
Rules for which data may reach which provider, enforced in code. Where compliance requires it, the app refuses to start with an unapproved provider configured.
Each feature gets a response-time target and a monthly spend budget. We pick the smallest model that passes the test set, cache stable answers and cap usage per customer.
AI features ship dark, reach a small group first and can be switched off per customer without a deploy. Every customization falls back to the standard behaviour when it fails.
An evaluation set built from your real inputs, scored on every prompt and model change. You see quality move as a number before customers notice anything.
Question answering over your documents with a vector store, re-indexed safely when the model changes and isolated per customer all the way into the search itself.
Product, wholly owned
The AI scribe ran on OpenAI. A HIPAA requirement moved the whole AI layer to AWS Bedrock: a new model, re-indexed document stores with a rollback path, strict input limits for real PDFs. The product’s core logic barely changed.
Read the seven lessonsProduct, wholly owned
AI added to a working CRM. The 130+ REST endpoints and 49 data models existed first; the MCP connector with 57 tools came on top, with OAuth 2.1 and tenant isolation enforced on every call.
Read the case studyPast performance, enterprise
Integration into large systems is backend work. CS2 spent three years (10/2022 to 10/2025) on Vodafone Kabel Deutschland’s field-operations REST backend under the GDPR: Java, Spring Boot, Kubernetes, AWS, Jenkins.
See past performanceWe read your architecture, find the integration points and classify the data each feature would touch. That decides which providers are even allowed.
The first feature goes into your real codebase, dark, against real inputs. Your team sees it working in context within weeks.
Provider abstraction, privacy rules in code, cost and latency budgets, and an evaluation set that runs on every change.
Flag on for a small group, then wider, with quality and spend tracked per feature. Your team gets the runbook and the kill switch.
Yes, and that’s the usual case. We start by reading your codebase and finding the right integration points, whether that’s a REST API, a message queue or the database layer. Our backend experience includes three years inside Vodafone Kabel Deutschland’s field-operations platform, which CS2 didn’t originate either.
With a provider abstraction, the switch itself is configuration. The work is around it: re-indexing any vector stores, re-running the evaluation set, and testing with real documents, because model limits differ. For Notermed the move was a short project, and the rest of the product kept running as before.
We classify the data each feature touches and write rules in code about which provider may receive it. For Notermed, the app cannot start if it is pointed at a provider outside the HIPAA boundary. Customer separation is enforced inside the AI search too, not only in the database.
It depends on volume, model and how much text each request carries, so we set a monthly budget per feature before launch rather than guess. Smaller models, caching and strict input limits usually bring the bill well below the first estimate. Usage is logged per feature and per customer so finance can see it.
With an evaluation set made from your real inputs and the outputs your experts would accept. Each prompt or model change is scored against it before it ships. After launch we track acceptance, edits and rollbacks per feature, so quality is a number you can watch.
You switch it off with a feature flag, per customer or globally, without a deploy. Customizations fall back to the standard behaviour on error, and the old index or model stays available for a rollback. We design for that before launch, because it always comes up.
Java and Spring Boot, Node.js and NestJS, Next.js and TypeScript on the application side; PostgreSQL and Redis for data; AWS for infrastructure. If your system speaks REST or publishes events, we can connect to it. For other stacks, we assess the integration points in the first call.