Kyiv AI Day 2027 Spring

More about resentation:
•⁠ ⁠"No requirements" is a false description of the problem. The knowledge exists; it is scattered, contradictory, and unindexed. Naming the real problem changes what you do about it - you are not writing documentation, you are reconciling sources.

•⁠ ⁠What AI is genuinely better at than you: reading everything without getting bored. Several hundred tickets, every test case, the commit history and the code, normalised into one structure. That is the part to delegate. Deciding what matters is not.

•⁠ ⁠The output that makes the difference is a graph, not a summary. Extract entities and the relations between them - feature to ticket to test case to code area to owner — and you can ask questions no individual document answers: what changes most often and is tested least, what has tests but no owner, what exists in the code and appears in no ticket at all.

•⁠ ⁠Assessment is a set of specific answerable questions, not an adjective. Risk concentration, coverage gaps, orphaned areas, and - the most valuable one - the places where two sources say different things, because that is where the undocumented decision is buried.

•⁠ ⁠Every claim carries a link back to the artefact it came from. Anything without one is a hypothesis and gets labelled as a hypothesis. This is the single rule that decides whether the assessment survives its first meeting.

•⁠ ⁠The failure mode is a confident answer that is wrong. The model summarises a component it half-understood in exactly the same tone it uses when it is right, and nothing in the output tells you which one you are reading. You need a second gate - sampling claims and checking them against the source by hand - and you need to say out loud how many you checked.

•⁠ ⁠What you hand over at the end. Not "the project is medium risk". A ranked list of specific unknowns, what each one would cost to resolve, and the two weeks of work that would remove the largest one. That is a document a manager can act on.

•⁠ ⁠And what this does not replace: the conversation with the three people who remember why. It turns three weeks of archaeology into three days, then tells you exactly which questions to ask them.