About the product
A SaaS platform where companies describe their sales process and train their sales reps on it. I built the AI part of the platform: an assistant that knows the company’s sales process, answers questions, teaches new hires and lets them practise conversations with a customer.
What I did
I owned the AI service from design, through implementation and model quality evaluation, to running it in production.
- A chat assistant that answers questions about the sales process.
- A learning mode in text and by voice, including audio explanations of each stage of the sales process.
- An AI trainer: a real-time voice conversation with a simulated customer at different difficulty levels.
- Quizzes, exam simulations and call script reviews.
- Progress summaries for managers.
Problems I solved
- Lag in voice conversations. A conversation with the trainer has to feel natural. I built and measured a new approach that makes the trainer respond more than twice as fast.
- An overly complex assistant. The first version of the assistant was needlessly complicated, and that complexity caused most of its errors. I simplified it, which improved the quality of the answers.
- Reports you can trust. In manager summaries, numbers and quotes have to be real. I designed them so the AI cannot make up a result or a quote.
- Two languages. The platform works in Polish and English, and adding another language is a predictable task.
- Security and data isolation. Each company’s data is kept separate from the others, and the assistant is protected against attempts to manipulate it through message content.
- Switching AI models. Moving to a newer AI model is quick and covered by tests, not a hunt through the whole codebase.
What I learned
- More AI agents is not better. A simpler solution gave measurably better results.
- Where trust matters, AI should write, not calculate. Numbers have to come from the data, not from the model.
- In voice solutions every link in the chain matters: from how the AI phrases its sentences to the choice of speech engine.
- I choose models based on tests with real cases, not on vendor claims.