← All case studies

E-COMMERCE / AI

AI System Optimisation, Observability & Quality Assurance

≈ +€60k revenue unlocked

3 new enterprise clients × ~€20k one-off project value each.

How we made an unreliable AI platform safe to sell, helping the team support larger clients and unlock new revenue.

LLMOpsLangSmithLiteLLMScalabilityE-commerce
Situation

An e-commerce startup with several AI services already running faced the challenge of scalability. The codebase wasn't standardised, redundant architectural layers were generating latency, and the lack of an observability and testing strategy made it difficult to guarantee the quality and flexibility needed as the product evolved.

Risk

The company could not confidently sell to larger clients while reliability, latency, and AI quality were hard to observe. Every new customer increased delivery and support risk.

Decision

Validate whether the existing platform could be stabilised and made observable enough to support higher-value customers, instead of rebuilding the product from scratch.

Intervention

We performed a full refactoring to standardise the codebase and removed redundant layers, optimising performance. We improved the management of tools like LiteLLM and LangSmith to gain full visibility into AI flows, refining prompts to improve their effectiveness. We then introduced a suite of unit and integration tests to lock in the correctness and quality of the exposed services.

Result

The business could finally sell the product with confidence. Reliability stopped being a sales risk: the team could see what was happening inside the AI flows, fix quality issues faster, and support larger clients. That stability helped unlock about €60k in new project revenue.

Economic value

≈ +€60k revenue unlocked

3 new enterprise clients × ~€20k one-off project value each.

Before / after

BeforeAI flows were opaque

AfterEvery interaction could be monitored and debugged

BeforeReliability limited sales

AfterLarger clients became manageable

BeforeQuality fixes were reactive

AfterIssues could be traced and fixed faster