From three lines of requirements to a ranked shortlist in minutes — with evidence, not guesses
A working proof of concept: free-text role requirements become an editable ideal profile, hard requirements are enforced by code (never by AI), candidates are ranked 0 100 by semantic match, and every screening question is answered with evidence from the CV or an explicit "no evidence". Built end-to-end in about a month.
Challenge
The problem pattern.
An assignment comes in. The requirements are two-three lines of free text or a posted job ad. Then someone shortlists from the pool: open CVs one by one, hunt for the degree, the past employer, judge whether the SQL experience is hands-on or just a word in a skills list. Meanwhile the clock runs — on supplier platforms, the first responder with the right candidate and a good pitch usually wins.
The judgement is rarely the hard part. The hard part is reading everything — and being first. Three things make naive AI fail here: hard requirements must be binary (a typo can't disqualify a candidate, semantics can't approve one), recruiters need specific questions answered per CV without reading them all, and the platforms often have no APIs.
Solution
What we build.
An agent-driven curator that owns the chain: requirement profile → selection → evidence → presentation. Three mechanisms most tools blur together are kept deliberately separate:
- LLM-generated requirement profile with a human in the loop — the agent drafts an ideal profile plus ~5 functional skills; the recruiter edits before anything runs. AI output is a starting point, never the verdict.
- Hard requirements as code — education, named employers and similar are boolean AND/OR filters applied deterministically. The UI shows the matching candidate count in real time, so you never build a filter that returns zero.
- Semantic match + evidence-based screening — candidates ranked 0–100; each screening question ("does the CV show hands-on SQL?") answered per candidate with reasoning or an explicit "no evidence". The agent is not allowed to guess.
A decided extension: the presentation layer — the candidate's CV re-weighted against the specific ad, plus a generated first-person pitch letter. A future trigger layer watches assignment platforms so the flow starts the moment a new ad is published.
Architecture
How it’s put together.
Candidate pool with document-AI extraction and indexing; LLM profile generation; a deterministic filter engine; a semantic matching engine with batched comparison views; a screening module with evidence requirements. Cloud-hosted on our Azure environment in the proof of concept; multi-tenant dashboard placement planned.
Questions about shortlisting
Can candidates be screened against requirements without an AI making the decision?
Yes. Hard requirements are checked by code, so nobody is excluded by a judgement call. Ranking is used only to order the candidates who already qualify, and every screening answer points at the place in the CV it came from, or states plainly that there is no evidence. A person still decides who to speak to.