IMPAD-02

Inspiration

Scaling technical teams is painful. The only way to verify a candidate's real knowledge is a deep 1-on-1 interview with a Senior Engineer. But Senior Engineers are expensive, busy, and unscalable. You cannot clone your best interviewer... or can you?
We realized that to fix hiring, we don't need another test platform. We need an autonomous digital interviewer. We wanted to build a system of independent AI agents that can conduct a technical deep-dive without a human present, objectively separating those who "know" from those who "memorized."

What it does

GlassBox is a desktop AI agent that audits a software engineer's workflow during technical assessments. Instead of just checking if unit tests pass, it tracks how the candidate works using two lenses:
Hard Skills (Workflow): It monitors the balance between Coding (IDE) and Researching (Browser), detecting efficient debugging patterns versus "doom-scrolling" .
Soft Skills (Clarity): It analyzes the candidate's spoken explanation using LLMs to score their logical coherence and terminology .

At the end of a session, it generates a session_log.json and a certified PDF report with a unique hash, proving the candidate's skills are authentic .

How we built it

The Brain (Cognitive Engine): We integrated OpenAI API to process the "think-aloud" protocols. We engineered strict system prompts to force the LLM to output structured JSON metrics (Coherence, Terminology) instead of generic text.

Challenges we ran into

Data Synchronization: Correlating the timestamp of a spoken word (Soft Skills) with the exact window state (Hard Skills) was tricky.

Accomplishments that we're proud of

Real-Time "X-Ray": We successfully built a pipeline where you can see the candidate's psychological state mapped onto a timeline. It’s not just data; it’s a story of their thought process.
Zero-Trust Architecture: We didn't just build a dashboard; we built a verification system. The generated PDF with a cryptographic hash proves that the skills are authentic and the report hasn't been tampered with.

What we learned

Process > Result: We confirmed our hypothesis that the path to a solution tells you more about an engineer than the solution itself.

What's next for IMPAD

IDE Integration: Moving from a desktop script to a native VS Code Extension for a seamless developer experience.
On-Chain Verification: Minting the session certificates as Soulbound Tokens (SBTs) so candidates can carry their verified skills across platforms (LinkedIn, Upwork).
Enterprise Scale: Adding role-based access control and integration with ATS (Applicant Tracking Systems) like Greenhouse or Lever.