← All projects
AI / Education

UNBLUFF

A mock placement interviewer that checks every claim on a resume against a target role and shows what the student can defend.

Role
Builder
Status
Working build · live Gemini grading
Category
AI / Education
Gemini APITypeScriptReactExpressFirestoreCloud Run
Your resume says you’re ready. Let’s prove it.
1,168
Backend tests
3
Interrogation levels per claim
Context

A student needs to know what to prepare, where they stand, what to improve and how to prepare. A resume full of claims says nothing about which of them the student can actually defend in an interview.

UNBLUFF turns the resume into a list of verifiable claims for a chosen role, then interrogates each one.

Approach

Every claim is probed on three levels: what you built, how it works, and why, including when not to use it. One follow-up is allowed per level when an answer is too vague. Each claim ends Defended, Shaky, Bluff or Honest gap.

The LLM only returns per-criterion booleans plus quotes. A criterion counts only if its quote is a verbatim substring of the student’s answer. Everything else is plain, unit-tested code, so the same answers always get the same score.

System

The model grades against criteria and quotes evidence. Code decides pass or fail, verdicts, readiness and priorities.

The model quotes, the code decides
  1. Resume to claims
    Claims mapped to the role’s weighted skills; unclaimed required skills become blind spots
  2. Three-level interrogationLLM
    What, how and why, with one follow-up per level
  3. Evidence guardCode
    A criterion counts only if its quote is verbatim in the answer
  4. Deterministic scoringCode
    Verdicts, proficiency, readiness, coverage and priorities
  5. Fix tasks and retest
    A targeted task per weak claim, then a delayed fresh-scenario retest
Deterministic codeLanguage model
Build
  • Report with readiness (0 to 100), role coverage, a resume heatmap and per-skill states, each backed by verbatim quotes
  • Top three priorities ranked by weight times one minus proficiency, with a root cause per weak claim and honest resume rewrites
  • Root-cause diagnosis from a per-skill prerequisite list, kept only if it matches the list exactly, at no extra LLM call
  • A 7-day plan, a 4-week roadmap preview and a delayed, interleaved retest that unlocks after two other claims
  • Gemini through the official Google Gen AI SDK with structured output and a fallback model; Firestore, Cloud Run and Secret Manager implemented for deployment
  • Strict request schemas, rate limits, prompt-injection defence and an accessible interface
Challenges
  • Keeping the model from deciding the score. A fooled model still cannot pass a criterion, because the evidence guard checks the quote against the real answer.
  • Proving improvement rather than memorisation. The retest is a new scenario, never a repeat, and records how many other claims came between.
  • Keeping cost down: report output is cached per claim, calls are limited to two at a time, and a quota cooldown keeps the free tier usable.
Result

A working build with live Gemini grading. The deterministic core has 100% line coverage, with 1,168 backend tests and 382 frontend tests passing at the latest run. Cloud deployment is implemented but not yet deployed.

Learnings

Let the model gather evidence and let code make the judgement. That is what makes a score repeatable and defensible.

Next project
PRYSM
AI / Compliance