RoleFit

5 tailored job artifacts in less than 50s with Gemini 3 Flash. A Validator Agent guarantees 0% fabrication, restoring trust to AI‑driven hiring with audited integrity.

Comment

RoleFit - AI Job Applications with Zero Hallucinations

RoleFit is an AI job application system built with Gemini 3 that refuses to hallucinate—every claim is validated against the user’s real experience before it’s shown or submitted.

Inspiration

After watching qualified candidates get auto-rejected by ATS filters—and seeing generic AI tools quietly invent experience that destroys recruiter trust—I built RoleFit to solve both problems: speed AND uncompromising accuracy.
One fabricated metric can end an interview before it starts. Job seekers need AI that's fast but also truthful and transparent. RoleFit delivers both.

What it does

  • Generates a complete ATS-ready application package in under 50 seconds
  • Produces five aligned artifacts using a single Gemini 3 context
  • Blocks hallucinations via a Validator Agent that verifies every claim against the source resume

Users get five coherent artifacts they can confidently submit:

  • Tailored Resume – ATS-optimized with verified experience only
  • Strategic Cover Letter – Narrative-driven, company-researched
  • LinkedIn Outreach Script – Personalized for hiring managers
  • Fit Analysis Score – Honest match percentage (1–100)
  • Gap Analysis Report – Transparent about missing qualifications

Every claim is grounded in the original resume. No hallucinations. No reputation inflation. Just truthful, strategic positioning.

How we built it

Two-stage pipeline

Stage 1: Structured generation with Gemini 3 Flash (JSON-constrained, long-context reasoning)
Gemini 3’s ability to reason over structured resume data and job requirements in a single coherent context is what enables consistent multi-artifact generation without drift.

Stage 2: Validator Agent
A fact-checking layer compares every claim (skills, titles, metrics, dates) against the source resume. If anything isn't grounded in the original document, it's rewritten or flagged. This prevents:

  • Skill inflation ("proficient" → "expert")
  • Title exaggeration ("developer" → "senior architect")
  • Fabricated metrics ("led team of 5" when the user never managed anyone)
  • Experience invention (adding roles that don't exist)

Architecture

  • Backend: FastAPI orchestrates the full pipeline as a unified "mega-call"
  • Frontend: React + Vite for responsive, mobile-first UI
  • AI: Gemini 3 Flash for generation + custom validation logic
  • Infrastructure: Google Cloud Run for auto-scaling deployment
  • Auth: Firebase for Google Sign-In and profile persistence

Despite the two-stage validation, total latency stays under 50 seconds through careful prompt optimization and parallel processing.

🏆 Verified 0% Hallucination Rate Across 110+ adversarial prompts designed to force experience inflation, the Validator Agent caught and blocked every fabricated claim.

Challenges we ran into

  1. Adversarial prompt resistance
    Getting Gemini 3 to stay honest under adversarial prompts (e.g., "make me look like a senior engineer") without ruining output quality was difficult. Users want aspirational language, but not exaggeration.
    Solution: Multi-layered prompt engineering with explicit fact-checking instructions, plus the Validator Agent as a safety net.

  2. Balancing strictness vs quality
    Too-strict validation made resumes feel robotic and undersold candidates. Too loose, and hallucinations crept back in.
    Solution: Iterative testing with 110+ adversarial cases to find the threshold where honesty meets persuasiveness.

  3. Reputation drift
    Entry-level candidates were being described as "senior experts" due to subtle AI tone inflation across multiple generation steps.
    Solution: Added seniority-level detection to the validator that flags title mismatches and adjusts language appropriately.

  4. Latency with validation
    Adding validation and extra reasoning steps initially doubled generation time to 90+ seconds.
    Solution: Optimized prompt structures, parallelized validation checks where possible, and streamlined the mega-call architecture to get back under 50 seconds.

  5. Mobile UX complexity
    Making professional job applications possible from a phone required rethinking the entire interface for thumb-friendly interactions.
    Solution: Mobile-first redesign with bottom-anchored controls, floating help buttons, and responsive layouts tested across multiple device sizes.

Accomplishments that we're proud of

  • 🏆 0% Hallucination Rate (Verified)
    Achieved a 0% hallucination rate across 110 adversarial test cases, including prompts designed to trick the AI into inventing experience (e.g., "make me look like a senior engineer with leadership experience"). The Validator caught and blocked every attempt.

  • ⚡ Sub-50-second generation
    Despite running both generation AND validation, maintained end-to-end processing under 50 seconds—faster than typical AI resume workflows that skip dedicated fact-checking.

  • 🎯 Eliminated reputation drift
    Fixed the subtle problem where junior candidates were described with senior-level language, ensuring outputs accurately reflect actual experience level.

  • 📱 Mobile-first production quality
    Shipped a fully functional mobile experience that makes serious job applications possible from a phone—complete with Firebase auth, multi-format downloads, and AI refinement.

  • 🔒 Trust-first architecture
    Built validation into the core system, not as an afterthought. Every output is fact-checked before users see it, helping ensure they never submit fabricated content.

What we learned

  • Trust isn’t about sounding impressive — it’s about verifiable accuracy. Users prefer honest positioning over exaggerated claims.
  • Gemini 3 performs best when constrained with structured inputs and explicit validation goals, especially in high-stakes domains.
  • Small edge cases (tone inflation, seniority drift) compound quickly in multi-step AI systems and require explicit safeguards.

What's next for RoleFit

  • Job application tracker
    Transform RoleFit from a one-off generator into home base for the entire job search. Track all generated applications with status updates (Applied / Interview / Rejected / Offer) and automated follow-up reminders.

  • Real-world callback measurement
    Collect anonymized data on interview requests and offers to validate whether our outputs improve callback rates beyond the 2–5% industry baseline. Use this feedback loop to continuously improve AI effectiveness.

  • University & career center partnerships
    Partner with universities and coding bootcamps to offer RoleFit as a "trusted AI application assistant" for students. Explore white-label deployments for career centers.

  • Interview prep integration
    Expand beyond applications into interview preparation, powered by Gemini 3:

    • Role-specific practice questions with feedback
    • Answer quality analysis and improvement suggestions
    • Mock interview simulations with realistic follow-up questions

  • Enhanced validation dashboard
    Show users exactly what the Validator caught and why, building transparency and trust. Let them see before/after comparisons of claims that were rewritten for accuracy.

  • Industry-specific optimization
    Build specialized validators for different industries (tech, healthcare, finance) that understand domain-specific terminology and credential requirements.

Built With

Share this project:

Updates