All work
Personal build
JD Matcher, a prototype for assessing Salesforce roles
A working prototype that compares a Salesforce role with a CV and shows which parts of the CV support its assessment.
Solo builder: product framing, prompt design, and application engineering · Ongoing, personal build · Salesforce hiring, career tooling
The problem
It is hard to judge whether a Salesforce role is worth applying for, and keyword-matching tools often make every profile look suitable. I wanted a faster way to assess a role using only evidence from the CV.
The intervention
I built a web tool that compares the role with the CV, produces a score and links strengths and gaps to the relevant CV passages. A separate set of rules checks the recommendation before it is shown.
The outcome
The prototype is available online and currently focuses on Salesforce roles. It is a working experiment, not a finished recruitment product.
The specifics
- Every score is evidence-linked: each matched requirement points to the specific CV text that supports it
- A deterministic calibration layer overrides the model when structured CV signals contradict its verdict
- Rate-limited and quota-gated per visitor, with JD and CV text sanitised before reaching the model
- Labelled openly as a prototype on the tool itself, not oversold as a finished product
How it works
- Browser-extension prototype reads a live public job page, requests confirmation before analysis, and returns the verdict beside the source role
- Claude tool-use call constrained to a fixed JSON schema: score clamped 0 to 100, enum role types, capped array lengths
- Structured candidate-signal extraction (certifications, clouds, capability keywords, Salesforce evidence level) drives rule-based recommendation calibration, independent of the model's raw output
- Gap count capped by recommendation tier, so a strong-fit verdict cannot hide unresolved blockers
- IP-based rate limiting and a daily AI quota on every request
Why you can trust it
- Same guardrail philosophy as the InsideBoard platform: the model proposes, deterministic code disposes
- Honesty by design: JD and CV input quality are graded and surfaced, not hidden, so the tool never claims more confidence than the input supports
- Built solo end to end: prompt design, schema, calibration rules, and the interface
Stack
- Next.js
- TypeScript
- Browser extension
- Anthropic Claude (tool use)
- Upstash Redis rate limiting
The prototype
See it live at judjit.comExperimental, not a finished product
Have a project like this?