SKILLFORGE

SkillForge simulates realistic technical interviews across AI/ML, digital project management, web development and digital culture: AI-generated questions that grow the question bank session after session, adaptive difficulty and category selection via Vectorize similarity search, LLM-as-judge scoring against an explicit rubric, a personalized revision plan and progress tracking over time — no account required.

PROJECT DETAILS

Objective

The goal was to turn generic interview prep into a personalized, adaptive coach: instead of a fixed question bank and a subjective grade, each session targets the candidate's weak spots and scores answers against an explicit, pre-defined rubric rather than a global impression.

Technologies used

The stack runs entirely on Cloudflare: a Hono API on Workers, D1 for sessions and question history, Vectorize for adaptive question selection by semantic similarity on weak points, and Pages for the vanilla TypeScript front end. Text generation — evaluation, follow-ups, questions, glossary entries, the revision plan — goes through OpenRouter; Workers AI is kept only for the embeddings that feed Vectorize. There is no user account: each browser gets a persistent anonymous client ID in localStorage, used to track progress and history without sign-up.

Challenges and solutions

The main challenge was scoring answers fairly and consistently: solved with an LLM-as-judge approach graded only against an explicit rubric of expected points defined in advance for each question, rather than a holistic impression, with neutral pedagogical feedback that never reads as a correction of the candidate's mistakes. Keeping the question bank fresh while staying relevant required generating new AI questions every session and feeding them back into Vectorize so adaptive selection keeps improving instead of cycling through a static set.