Interview Preparation · NAB Melbourne · AI Scientist
Intent Recognition
Working out what the user actually wants, and turning it into something the system can execute. Two pages: the discipline from the ground up, then exactly how it is implemented in the Mines Financial project — including the parts that are missing.
Why this topic matters here
It is the first line of the job description's responsibilities, phrased as something else.
The NAB advert asks for someone to build agent reasoning loops, tool use interfaces, memory systems, context management, and orchestration logic. Intent recognition is the hinge between the first two: it is the step that turns a sentence into a decision about which tool to call. Every one of those five things depends on getting it right.
It also connects directly to the two topics the advert leans on hardest. Its failure mode is a guardrails problem — a wrong intent is not a bad paragraph, it is a wrong action — and the way you know whether it works at all is an evaluation problem, with its own precision, recall and confusion matrix.
The two pages
Read them in order. The first sets up the vocabulary the second one uses.
The concept — ground up, nothing assumed
What intent recognition is, why an agent cannot function without it, and how to design one properly: pick intents from business actions rather than topics, split them only when what happens next differs, define each by its boundaries with its neighbours. Then the three ways to build it — a trained classifier, an LLM asked to classify, or no classifier at all with intent implicit in tool selection.
The back half is measurement: why accuracy is misleading, what precision and recall each tell you, how a confusion matrix tells you which specific pair of intents to fix, the four layers from intent through to task success, and the metric that outranks all of them for an agent that can act — wrong action rate.
- Includes — the granularity test, include/exclude definitions, hierarchical routing, unknown and ambiguous intents, multi-intent as task planning, and how to build a test set that is not flattering
- Ends with — the whole discipline as one chain, and a glossary of every term used
In this project — what actually runs
Mines Financial has no file called intent_classifier.py. What it has is one node returning four independent booleans, and a set of deterministic edges reading them. This page walks every decision point in execution order — from history compression, through regex slot extraction and follow-up resolution, to the router itself, the clarification interrupt, per-company query decomposition, and the two capped retry loops.
It also covers the parts that are absent and why that is defensible: no confidence on any routing decision, no explicit check for an unsupported company, no loop counter on clarification. And the honest footnote about the routing evaluation partly grading its own test harness.
- Includes — why slots are regex rather than a model, why one decision is made deliberately late, and why almost nothing here is a free tool choice
- Ends with — a ninety-second spoken answer, and the gap to close on
The one thread through both
Part one gives you the chain. Part two shows the same chain with this project's names on it.
| The general step | In Mines Financial | Made by |
|---|---|---|
| What does the user want? | Four flags from retrieve_decision | One LLM call, structured output |
| What information does it need? | Companies and fiscal years | Regex — closed vocabulary |
| Is anything missing? | needs_clarification → graph interrupt | LLM flag, then human |
| Which workflow? | Retrieval and/or news, in parallel | Deterministic edge |
| How many searches? | One per company named | Second LLM call, decomposition |
| Which tool? | Graph edges everywhere but the calculator | Code, except one ReAct agent |
| Did it work? | Groundedness count → confidence label | Threshold on unsupported claims |
Reading the table left to right is the concepts page. Reading it right to left is the project page. Being able to move in both directions — general principle to specific implementation and back — is what separates an answer that sounds researched from one that sounds lived.
Questions to be ready for
Where each answer lives.
| If they ask | Go to | Lead with |
|---|---|---|
| "How does your agent decide what to do?" | Part 2 | Four flags, not a label — then why that allows multi-intent |
| "How would you design intents from scratch?" | Part 1 | Business actions, not topics; split only when execution differs |
| "How do you know the routing is any good?" | Part 1 then 2 | Confusion matrix over accuracy; then the 8 routing cases |
| "What happens when the query is ambiguous?" | Both | Confidence plus a clarification policy; here, a real graph interrupt |
| "What if the user asks two things at once?" | Part 2 | Two axes handled, one not — and why that one is harmless here |
| "What would you do differently?" | Part 2 | Routing confidence, before ever adding a tool that writes |