qado.ai IDP: RAG and AI Agents for Procurement

Overview

This is my Interdisciplinary Project (IDP), the part of the TUM master’s degree that applies computer science in another discipline. It ran with the Chair of Financial Accounting at the TUM School of Management, in partnership with qado.ai, a Munich startup that builds AI tools for analysing contract and procurement data for corporate customers. The project ran full time for three months, inside the startup’s product and agile process.

In strategic purchasing, good decisions depend on what contracts actually say: prices, terms, obligations and renewal conditions, spread across long, unstructured documents. The goal was a production-oriented component of qado.ai’s application that makes that knowledge searchable and usable, first for people and then for software agents acting on it.

The retrieval pipeline

The core was a retrieval-augmented generation pipeline built as separate, swappable stages. Contract documents are parsed, the relevant content is extracted and structured, and it is embedded in a vector space. At question time, retrieval finds the passages that matter and passes them to a language model to work with.

Retrieval on contracts is harder than on general text. The right answer can depend on an exact clause name, which favours keyword search, or on meaning, which favours vector search. It can also depend on how documents relate, such as a framework agreement and its amendments, which calls for graph traversal. The project therefore evaluated these methods side by side and in combination, together with filtering before and after retrieval and re-ranking of the candidates.

Measuring retrieval quality

Those choices should be made on evidence, not intuition. Part of the project was a benchmark for retrieval quality on this kind of data, used to compare strategies and optimise the pipeline step by step. This is the same discipline as my cross-lingual RAG study: fix the questions, know the right answers, and measure what each change actually does.

Agents for purchasing tasks

On top of retrieval, the project prototyped simple AI agents for steps in the purchasing process, such as supporting communication with suppliers or identifying levers for a negotiation. Each agent combines retrieval, language-model reasoning and explicit decision logic, so it can analyse contract content and suggest a course of action. The architecture was kept modular and extensible, so new agents and data sources can be added without rebuilding the pipeline.

Business context and timeline

The project connected machine learning, information retrieval and software engineering with business questions in purchasing and financial management. The companion module, Investment and Financial Management, provided the corporate-finance grounding to judge the system by what matters to the business, such as cost and efficiency, not only by technical metrics.

The work ran in three phases. The first month covered the data sources, systems and requirements, the architecture, and first data preparation. The second built the RAG pipeline with embeddings and retrieval, followed by first evaluations and iterative improvement. The third optimised the retrieval strategies, completed the benchmarking, and developed and integrated the agent components, followed by documentation and a final presentation.