Duster
Sunset mountains seen through a train window

Rejections never come with feedback.

See the weakness that keeps costing the offer.

Dump the interview the way you remember it. Duster finds the pattern across rejections, then drills the answer until the next one is stronger.

Log a debrief See the features

Features

Everything after the interview, in one place.

She leaves the call, gets rejected or ghosted, and never hears why. Duster keeps the interview the way she remembers it and turns the repeat mistakes into practice.

Debrief

Log the company, the role, and how it ended, then dump what they asked and what she said. The formatting can be messy.

  • Outcome is rejected, ghosted, offer, or unknown.
  • Q and A blocks can use colons, dashes, or markdown.
  • Each answer is tagged as soon as the debrief is saved.

Patterns

Weakness tags are counted across every interview, not just the last one.

  • Shown as “no_metrics — 7x in 4 sessions.”
  • An arrow marks whether that tag is rising or falling.
  • A drill streak sits next to the outcome badges.

Drills

Five practice questions written for her target role. She answers, then sees the gap.

  • Score from 0 to 10, plus the tags on that answer.
  • A stronger rewrite of her words, not a generic sample.
  • Focus mode aims the questions at her top two tags.

Voice notes

If typing the dump is too much, she can drop the voice memo from the lobby.

  • .m4a and .mp3 only, transcribed on this machine.
  • The text prefills the dump. The audio file is deleted.
  • The recording never leaves the server.

Memory

One Backboard assistant keeps a thread of her practice, so the next session is not a blank page.

  • Each drill logs the question, the score, and the top tags.
  • The patterns page reads that thread back as a short summary.
  • Tagging an interview does not write into that history.

Open models only

Every tag comes from rules or an open-weight model. A bad model reply falls back to the rules instead of crashing the page.

  • Heuristic: length, missing numbers, filler, missing structure. No key.
  • Ollama: local qwen2.5:7b on her machine.
  • Backboard: one key, open weights only. Closed providers are refused.
  • Tinker: the LoRA fine-tune, same tag contract as the others.

Pricing

One plan. It is free.

Duster is built for one person. There is no seat fee and no upgrade. Optional model keys are yours to add; the coach itself stays at zero.

Current

Free

$0 for one job hunter

The whole loop, with no account and no card.

The coach

  • Unlimited debriefs, stored in a local SQLite file
  • Pattern counts, trend arrows, outcome badges, and a drill streak
  • Five-question drills with a score and a rewritten answer
  • Focus mode aimed at her top two weakness tags
  • Settings for target role, domain, and self-reported weak areas

On this machine

  • Heuristic tagging with no API key
  • Voice notes transcribed locally, then deleted
Start a debrief

Still $0

Optional models

Turning a model on does not change the plan. Duster never calls a closed API.

Included either way

  • The site runs on the heuristic if no keys are set
  • A failed model response falls back to those same rules
  • Audio stays on the server that received the upload

Only if you add a key

  • Backboard tags answers and remembers drill history
  • Ollama runs qwen2.5:7b on your own machine
  • Tinker serves the fine-tuned checkpoint
Set role and weak areas