Candidate model
- Structured schema: history, skills, target roles, evidence stories, voice
- LLM-assisted intake from existing resume + LinkedIn export
- Versioned per-edit; every artefact tags the version it drew from
Group 11
· Capstone · Spring 2026
AI-Powered Job Search — the career stack, consolidated.
A unified job-search assistant that consolidates the fragmented career stack — resume tailoring, job-board ingest, outreach drafting, interview prep — into a single workflow grounded in the user's actual history, with clear ownership of every piece of generated text.
IWhat we built
What problem this solves.
Looking for a job is a coordination problem dressed as a content problem. The candidate ends up running five tools at once — resume tailoring in one tab, job-board searches in another, outreach drafts on a third, interview prep on a fourth — and the consistency between them is the candidate's burden to maintain. Most existing AI tools optimise one tile of that grid while making the rest worse.
Group 11 built the assistant for the candidate carrying that whole grid. The product is not another resume generator. It is a single workflow whose internal model of the candidate — history, target roles, voice — stays consistent across every artefact it produces.
IIHow it works
The system, end to end.
The system's spine is a structured candidate model: history, skills, target roles, evidence stories, voice. Every artefact — a tailored resume bullet, a cold-outreach draft, an interview-prep card — is generated from that single model. Edits in any surface propagate back to the candidate model, so a fix in one place lands everywhere it matters.
Job-board ingest is treated as retrieval, not browsing. The system pulls structured postings from supported boards, normalises titles and required skills against the candidate model, and surfaces matches with the gap explicit ("matches on 7 of 9 required skills; missing: Kubernetes, GraphQL"). Outreach drafts and resume tailoring start from the gap and the match together — not from the user staring at a blank textarea.
Ingest
Candidate profile
history · skills · voice
Ingest
Job-board ingest
multi-source crawl
Model
Candidate model
versioned per edit
Model
Artefact generators
resume / outreach / interview-prep
Storage
Candidate versions
every edit tagged
Surface
Propagation
drift flag + regenerate
Ingest Model Storage Surface
Feedback
IIIThe stack
What it's built on.
| Candidate model | Structured schema: history, skills, target roles, evidence stories, voice |
|---|---|
| Job-board ingest | Adapters for supported boards, normalised to a posting schema |
| Artefact generators | Resume tailoring against (candidate, posting) pair |
| Propagation | Edit-tracking surface flags drift between artefact and model |
Job-board ingest
- Adapters for supported boards, normalised to a posting schema
- Title + required-skills extraction with calibrated confidence
- De-duplication across boards via canonical (company, role) tuples
Artefact generators
- Resume tailoring against (candidate, posting) pair
- Outreach drafting in the candidate's stated voice
- Interview-prep cards anchored to evidence stories
Propagation
- Edit-tracking surface flags drift between artefact and model
- Per-artefact "based on" pin showing the source model version
- Stale artefacts marked for regenerate when the model changes
IVDeliverables
What the team shipped.
Source repository GitHub · code, tests, README
Demo video Capstone day · screen recording, 4–6 min
Write-up PDF Final brief · methods, evaluation, reflection
Slide deck Capstone presentation · 10 slides
VWhat sets it apart
What sets this capstone apart.
Resume, outreach, prep — same source.
Every artefact starts from the same structured candidate model. The resume and the outreach draft can't disagree about what the candidate did last summer because they both read the same field.
Show what's missing, not just what fits.
Most job-match tools surface the score and hide the why. This one surfaces the matched skills and the missing ones side-by-side. The candidate decides whether to close the gap or skip the posting.
The candidate model is the truth.
An edit to a resume bullet updates the underlying candidate model; the outreach drafts and interview prep that drew on that fact update with it. There is no drift between artefacts because there is no second source.
VIIInstructor note
How this project landed.
Job-search products are easy to start and hard to finish. The proposal showed the usual symptom: a long feature list, with every feature gestured at and none owned. The first review cut to the consistency thesis — one candidate model, every artefact derived from it, every edit propagated — and the rest of the work followed from that.
The capstone shipped a working assistant that does the resume / outreach / prep loop with internal consistency the comparable consumer tools do not enforce. The team defended the model-as-truth contract against the obvious shortcut ("just regenerate everything") and held the line.