How it works
How the diagnostic finds a gap you did not know you had.
Assessmentr collects evidence through a voice conversation, maps it onto AI/ML concepts, and returns one gap with the answer it came from. This page walks the whole thing through on a sample session.
The sample session, in full
Illustrative sampleThis walkthrough uses a synthetic session written to show how the diagnostic behaves. It is not a real user result.
The full sample session, start to finish.
Veda
Walk me through what happens to memory on a serving box while a transformer generates a long response.
Step 01
Veda chooses the concept, not you.
The opening question is selected from the benchmark and from what the diagnostic has no evidence for yet — not from a topic you nominated.
Candidate
So the KV cache is doing the work there. You keep the keys and values from the tokens you've already processed, so each new token only attends against cached state instead of recomputing the whole prefix. That's what takes you from re-running attention over the full sequence every step down to one step's worth of work.
Step 02
Correct, confident — and not an answer.
This explains why caching makes generation fast. The question was about memory. Most self-review stops here, because nothing said was wrong.
Veda
That's the latency side, and it holds. I asked about memory though — as the sequence gets longer, what sets how large that cache gets, and which part of the architecture controls it?
Candidate
It grows with sequence length, and with the number of layers since each one keeps its own. Beyond that I would have to look it up — I am not sure what else you would change to control it.
Step 03
You never asked about grouped-query attention. The diagnostic did.
Veda notices the answer covered latency and left memory untouched, and follows it into the prerequisite underneath: what actually sets the size of that cache.
LLM Engineer
Preliminary picture formingKV cache purpose
Explained why cached keys and values remove repeated work during generation.
KV cache memory growth
Recalled sequence length and layer count, but not the head dimension that memory also scales with.
Quantization tradeoffs
No evidence collected in this area yet.
Evidence basis
It grows with sequence length, and with the number of layers since each one keeps its own.
The gap comes from this answer and nothing else. Confidence stays preliminary because one follow-up is thin evidence for a concept this broad.
Study this next
Grouped-query attention and KV cache size
Work out how cache size scales with heads and head dimension, then compare multi-head, grouped-query, and multi-query attention on the memory they hold at long context.
Step 04
One gap, traced back to what was said.
The cache also scales with the number of attention heads — which is exactly the term grouped-query and multi-query attention shrink. That is the concept the session surfaced, and the only one it recommends.
What the evidence attaches to
Answers become evidence against a concept map.
Every AI/ML area is broken into concepts the diagnostic can actually assess through conversation. Evidence attaches to the concept it came from, which is why a gap can always be traced back to something you said.
01
Domain
Large language models
The broad AI/ML area a benchmark draws from.
02
Topic
Inference and serving
A cluster of concepts that tend to be examined together.
03
Concept
KV cache memory growth
The level you see. Assessable in conversation, explainable, and actionable on its own.
Three numbers that are not the same number
Mastery, coverage, and confidence are tracked separately.
Mastery
How the evidence collected so far reads for a concept. Shown as a label, never as a percentage.
Coverage
How much of the benchmark the session actually reached. Low coverage caps how strongly anything can be claimed.
Confidence
How much the result should be trusted yet. One answer on a broad concept stays preliminary, however clean the answer was.
Questions people ask before starting
- Is the walkthrough on this page a real session?
- No. It is a synthetic sample written to show how the diagnostic behaves, and it is labelled as such wherever it appears. Real results are private to the person who ran the session unless they choose to share one.
- Do I have to speak?
- Voice is the primary channel because explaining out loud exposes gaps that multiple choice hides. A text fallback is available inside the session if you cannot speak.
- How long does a session take?
- The diagnostic is adaptive, so it ends when it has enough evidence rather than after a fixed count. Most first sessions run around fifteen minutes.
- Why is there only one recommendation?
- Because a list of twelve things to study is the same as no plan. The diagnostic ranks candidate gaps and returns the one with the most leverage for your benchmark.
- What if I disagree with the gap?
- Say so. Every claim shows the answer it came from, and the summary lets you reject it. That feedback is recorded against the session.
- Which roles does this cover?
- AI/ML engineering only — ML Engineer, LLM Engineer, Applied AI Engineer, AI Infrastructure Engineer, and MLOps Engineer. It is not a general software diagnostic.
See the gap map this session produced
The same sample result, on the page a real session would hand you at the end.
Open the sample gap mapRun it on yourself.
One session. One gap. One thing to study tonight. Free during beta.
Start diagnostic