Under the Hood

The pipeline, in order

Submitting a case runs a short, real sequence, not one black-box call:

1. 01 — Photo or PDF uploads get read into plain text by a vision model, if that's how you submitted.
2. 02 — Your resume text is scored: keyword match, weak verbs, passive voice, buzzwords, length. Zero AI here — pure computation over the actual text.
3. 03 — The roast is generated in character, grounded in what the score just found.
4. 04 — The rewrite runs, then checks its own output and retries if needed — one of the three genuinely agentic steps, explained below.
5. 05 — It then re-scores its own rewrite with the same computed check from step 02, and sharpens once more if job-posting keywords are still truthfully missing — the second agentic step.
6. 06 — A cover letter draft runs the same authenticity check-and-retry loop, if you ask for one.
7. 07 — If you start a practice interview or negotiation, each next question is generated by reading your actual last answer and deciding whether to press deeper or move on — the third agentic step.

Which model does what

Each task gets a model actually suited to it, not one model forced to do everything:

• Reading an upload — a vision-capable model, since it has to actually see the page.
• The roast — a model chosen for personality and wit, since a persona that hedges into blandness isn't worth reading.
• The rewrite and cover letter — a smaller, more literal model, since the job is mechanical (turn a weak bullet into a strong one) not creative.
• Practice interviews & negotiation — the same model as the roast, since staying in character and judging whether an answer actually held up both need real personality and judgment.
• The score — no model at all. Keyword matching, verb detection, and length checks are plain code, which is why they're instant and reproducible.

The three genuinely agentic parts

Most of what's above is a single request in, a single response out — that's “AI-powered,” not agentic, and we'd rather say that plainly than borrow a word that doesn't apply. Three parts of the pipeline actually earn it.

The self-check-and-retry loop.The rewrite and the cover letter draft each check their own output against a deterministic rule (does this contain AI-cliché phrasing like “leverage” or “spearheaded”?), and if it fails, the system decides on its own to try again — feeding the model back exactly which phrases it used and got caught on, up to three attempts, before ever showing you a result.

The score-improvement agent.After the rewrite, the system re-runs the exact same computed keyword-match check from step 02 against its own output — zero AI cost, same math the score card shows you. If specific job-posting keywords are still missing, it decides on its own to sharpen the wording once more, but only rephrasing a bullet toward a keyword where that bullet's own content already genuinely supports it. It never adds a skill or responsibility the original bullet didn't already contain, and it tells you plainly when a keyword just doesn't apply rather than pretending it fixed everything.

The adaptive interview agent.Practicing an interview or a negotiation isn't a fixed script. After every answer you give, the model reads it and decides for itself — press deeper on the same topic because the answer was vague or dodged the question, or move to a genuinely new area because it already held up. That decision is made fresh each turn, for up to 5 turns, then it renders a real verdict on the whole conversation.

That's the actual definition doing the work here: the system observes its own output (or your answer), makes a decision based on that observation, and acts again without being asked to — a real loop, not a label. Run any of them yourself: a rewrite that took more than one pass says so, the before/after score is shown plainly, and an interview session visibly changes direction based on what you actually said.

Why we're being this specific about it

The rest of this site calls out other resume tools for repeating a scary, unsourced statistic to sell an upgrade. It would be a strange kind of hypocrisy to turn around and call four single-shot API calls “agentic AI” because the phrase is trending. So this page says exactly which part earns that word, and leaves the rest as what it honestly is.