Sign inStart
Algorithms · System design · Behavioral

They said no. You still don’t know why.

Gadfly is a voice tutor for technical interviews. You think out loud; it asks the next question instead of handing you the answer, and it remembers what broke last time.Gadfly is a voice tutor for technical interviews. You think out loud; it asks the next question instead of handing you the answer.

It listens while you think, then names the one thing that went wrong — in your own words, with the minute it was said. Tomorrow’s plan changes because of it.It names the one thing that went wrong — in your own words, with the minute it was said.

Start a sessionSee a real session ↓

Free while it is early. Your microphone never leaves your machine.

The end of a session: the one thing that went wrong, quoted in your own words
A real wrap-up. The badge means the same thing broke twice in a row — that is why it is on tomorrow’s list.
Why it stopped working

Three hundred problems. All of them in your head.

You can solve — you proved that three hundred times. Nobody ever heard you do it, and the room scores only what it hears.You can solve. Nobody ever heard you do it, and the room scores only what it hears.

“the feedback primarily boiled down to a very repetitive you need to go deeper”

Blind · on a $100 mock with a Google L7

“My brain is straight mush. Concepts blending together, learn 1 new thing, push out another”

Blind · thread “Anyone else’s brain fried?”
1 in 5

That is how many engineers perform the same way twice.

The same person can pass one interview and fail another on the same day.

interviewing.io, across thousands of interviews

Three rounds

Each round breaks you a different way.

Not one of them is testing what you know.

Algorithms

You solve it. They still say no.

You get the answer right, but you think in silence. The room scores what it hears, and it heard nothing.You get the answer right, but you think in silence. The room heard nothing.

What happens instead

It will not take your code until you name the invariant — the one thing that stays true at every step. And there is a button that says “Tell me the answer”: press it and the answer does not come, the question you were dodging does.No code until you name the invariant. Press “Tell me the answer” and you get a better question instead.

System design

You know the parts. You cannot put them together.

You can name a cache, a queue, a shard. In the room you draw boxes before you ask what the system must do — and you talk at the interviewer instead of with them.You draw boxes before you ask what the system must do, and you talk at the interviewer.

What happens instead

The model answer stays locked until you say yours out loud. You draw on a board; it draws back what it heard, so you see how you came across.The model answer stays locked until you say yours out loud.

Behavioral

Your story is good. It is scored as weak.

You tell a real story from your work. They hear “we decided” where they needed “I decided”, and a result with no number in it.They hear “we decided” where they needed “I decided”, and a result with no number.

What happens instead

One hard stop: it does not move on until you name a decision as yours. A cell on your map turns solid only when a session actually heard it.It does not move on until you name a decision as yours.

Everything above is in the app today. Nothing here is a plan.

A session, start to finish

This is what one evening leaves behind.

You talk it through. These three screens are what is there in the morning.

The tutor answers a request for the answer with a question
You talk. It will not answer.Bottom right: “Tell me the answer” is a real button. Press it and the answer does not come — a better question does. That refusal is the whole product.“Tell me the answer” is a real button. Press it and you get a better question, not the answer.
Six axes scored against your own rating of the session
It scores what it heard.Computed by rules over the session journal — not written by a model. The gap is between your number and its.Computed by rules over the journal. The gap is between your number and its.
Tomorrow’s queue, with every item traced to the session it came from
Tomorrow is different.Every line says why it is there: which pattern, which drill, which day of the repeat. Nothing lands on this list that you cannot trace back.Every line says why it is there — the pattern, the drill, the day of the repeat.

Every screenshot on this page came from a real session. None was drawn.

Why this teaches

Four things that are not features. They are the method.

None of this is new pedagogy — it is the oldest teaching there is, wired into a product that will not break its own rules when you push.Not new pedagogy. The oldest teaching there is, wired in so it holds when you push.

It refuses

The Socratic part, and the only one people argue with. Ask for the answer and you get the question you were avoiding. A tutor who caves is a search engine with manners — and you already have one of those.Ask for the answer, get the question you were avoiding. A tutor who caves is a search engine.

You say it out loud

Explaining is not the same skill as knowing, and only one of them is scored in the room. Thinking silently feels like studying; saying it badly out loud is the thing that actually moves.Explaining is a different skill from knowing — and it is the one the room scores.

It comes back on a schedule

What you got wrong returns after a day, then three, then a week, then two. Spaced repetition, applied to reasoning instead of vocabulary — because the slip you fixed on Monday is the slip you make again on Friday.Wrong today comes back in a day, then three, then a week, then two.

Nothing closes on one good day

A mistake clears after three clean runs in a row, not one. One good evening is luck; three is a change — and the difference matters more the closer the interview gets.A mistake clears after three clean runs, not one. One good evening is luck.

Every other tool on this market sells you another run at another problem, then forgets you. The method above only works if something remembers — which is the entire reason this thing has a memory at all.Everything else sells another run and forgets you. This method only works with a memory.

What eight weeks look like

Eight weeks. Same hours. Different engineer.

The hours are still yours to put in. What changes is where they go.

Week 1

You find out what actually breaks.

It is usually not what you thought. Most people guess “I need more problems”. The journal says something narrower, and quotes you saying it.It is usually not what you thought. The journal says something narrower, in your words.

Week 4

The same slip stops coming back.

Three clean runs in a row and it clears from your plan. Then the next one surfaces, and it is a smaller one.Three clean runs and it clears. The next one that surfaces is smaller.

Week 8

The room is no longer a surprise.

By then you have said your reasoning out loud more times than you can count. The interview is one more of those — that is the whole trick, and there is no shortcut in it.By then you have said your reasoning out loud more times than you can count.

No score, no percentage, no “you are 78% ready”. Those are guesses. This is a journal.

It also stops asking you to track any of this. The map knows which patterns are solid and which keep slipping, and the plan comes from that.

What it costs

A mock with a real engineer runs $100–$300 an hour. An evening here costs 66 cents.A real mock is $100–$300 an hour. An evening here costs 66 cents.

The split is why it is cheap: the thinking runs in the cloud, the listening and the speaking run on your machine. Voice is the expensive part everywhere else, and here it costs nothing to run.

66¢a session
What an evening costs
That is the whole bill for one session: the model doing the thinking. Speech in and speech out add nothing — they never leave your machine.
Your keyyour bill
Or run it yourself
Point it at a hosted model with your own API key and pay that bill directly. Most of the prompt is cached, so the price stops growing as the session gets longer, and you set a daily ceiling in dollars.
$0fully local
Or nothing leaves the room
Run the model on your own GPU too. Slower to answer, and it still teaches — and then not one word of the session goes anywhere.

The ceiling is the only thing that ever locks: past the amount you set, live calls stop and the tutor says so out loud instead of quietly turning into a template machine. Speech recognition and the speaking voice always run on your machine — there is nowhere for the audio to go.Past the ceiling you set, live calls stop and the tutor says so. Speech never leaves your machine.

Straight answers

The questions you are already asking.

AI mocks are free now. Why pay?

Because free ones forget you. Every product on this market sells another run at another problem; none of them remembers the last one. The value here is not the session — it is the eight weeks of sessions that know about each other.

I only need help with algorithms.

Behavioral is 30–40% of the 2026 loop, and strong coders lose offers there more often than on code. You can turn the other tracks off — but turn them off knowing that.

Does my voice go to your servers?

No. Speech recognition and the speaking voice both run on your machine, and there is nowhere for the audio to go. What can leave is the text of the conversation, and only if you point the tutor at a hosted model — run it on a local one and nothing leaves at all.

What if I already have my own problem list?

Paste it. Those problems have no local tests, so there is no Run button — you check the answer where it came from, and the session verifies it the way an interview does: by walking your code out loud.

Who is this not for?

Anyone who wants the answer. He will not give it — that is the entire product. If you have twenty minutes and want a hint, this is the wrong tool.

The work is still yours. Start tonight.

No card, no plan. Your voice stays on your machine.

Start a session