AI Project Management Assistant

Context

A conceptual demo existed internally, with the potential to become a product that reduced manual documentation and project tracking work.

Problem

Project meetings generate decisions, tasks, owners and knowledge that afterwards have to be organized by hand — and often simply get lost.

Constraints

• A new domain with no predefined path: the technical solution had to be researched, not just implemented.
• An internal MVP: enough scope to prove the idea, not to ship a commercial product.

The work

• Turned the conceptual demo into a working web MVP and helped define the technical solution.
• Designed and integrated AI flows: LLMs, prompt engineering, RAG, embeddings, vector databases, semantic search and contextual memory.
• Used meeting audio and documents as sources of information through speech-to-text and document processing.

Decisions

• Orchestrate the AI flows with tooling (Dify, n8n) instead of hard-coding each step, to keep iteration cheap.
• Retrieval over raw context: embeddings and semantic search instead of pushing everything into the prompt.
• Keep the human in the loop — the system proposes, the team decides.

Result

A functional base for a product able to turn meetings into useful operational knowledge — an internal MVP with commercial potential, not a finished product.

Learnings

• Working with AI is mostly product work: the hard part is deciding what the model should be trusted with.
• Autonomy in an undefined problem is a skill of its own — research, decide, build, discard.