SkillForge

An AI-powered adaptive onboarding engine that analyzes your resume against any job description, mathematically calculates skill gaps, and generates a personalized, gamified learning roadmap.

Comment

🔥 Inspiration

Every year, the tech industry loses billions forcing talent through generic, one-size-fits-all onboarding. Experienced hires waste up to 40% of their time on redundant training modules, while freshers get overwhelmed because they don't know exactly which micro-skills they are missing. We set out to fix this broken system — not with another course catalog, but with an intelligent, AI-native engine that adapts to each individual learner.

🧠 What It Does

SkillForge is a unified AI ecosystem that takes a candidate's Resume (PDF) and a target Job Description, then:

  1. Extracts skills using Google Gemini 2.5 Flash with structured JSON outputs validated by Pydantic schemas.
  2. Calculates the precise skill gap using a Semantic Embedding Pipeline (text-embedding-004) that computes Cosine Similarity — so the engine understands that "React.js" satisfies "Frontend Development."
  3. Constructs a Directed Acyclic Graph (DAG) using NetworkX to logically sort foundational prerequisites before advanced modules.
  4. Generates an adaptive, gamified roadmap rendered as a 60fps interactive React-Flow topological map with color-coded severity nodes (Critical / Partial / Near-Competent).
  5. Embeds the top 6 curated YouTube masterclasses per skill gap directly into the dashboard using a Zero-Auth Protobuf Proxy, completely bypassing restricted API quotas.
  6. Exports a personalized 6-page AI curriculum PDF with weekly study schedules via jsPDF.
  7. Hosts a real-time Multiplayer Quiz Arena where users challenge peers in host-controlled waiting rooms, ending in a post-match AI analytics dashboard that explains every incorrect answer.
  8. Includes a NotebookLM-Style AI Mentor — paste any youtube video URL,article or doc site, and Gemini instantly synthesizes the lecture into condensed, actionable learning takeaways.

🏗️ How We Built It

Backend Architecture:

  • Built an asynchronous Python server on FastAPI with Pydantic v2 schemas for strict request/response validation.
  • Engineered a Semantic Embedding Pipeline using Google's text-embedding-004 API to vectorize and compare skills via mathematical Cosine Similarity, replacing heavy local PyTorch models to stay under Render's 512MB RAM limit.
  • Constructed skill prerequisite graphs using NetworkX DAGs, then ran topological sorting to generate logically ordered learning paths.
  • Invented a Dual-Pronged Token Load Balancer that dynamically cascades from gemini-2.5-flash to gemini-2.0-flash if Google's rate limits are hit, guaranteeing 100% API uptime.
  • Built a Zero-Auth Protobuf YouTube Proxy using youtube-search-python with a custom httpx compatibility patch to scrape and rank videos by view count and channel reputation — no API key required.

Frontend Architecture:

  • Built a multi-page React SPA with React Router and Clerk authentication.
  • Rendered the adaptive roadmap using React Flow with animated edges, custom node styles, and real-time progress tracking with confetti celebrations on module completion.
  • Designed a responsive, glassmorphic dark-mode UI with dynamic micro-animations and hover effects.

Real-Time Features:

  • Engineered an asynchronous Multiplayer Quiz Arena with host-controlled lobbies, live scoreboards, and individual answer tracking.
  • Built a Post-Match AI Analytics Dashboard that maps each player's historical selections against AI ground-truth explanations.
  • Integrated a NotebookLM-Style AI Mentor pane that uses Gemini's native Part.from_uri to directly ingest YouTube video URIs and produce structured learning summaries.

🧱 Challenges We Faced

  1. Gemini API Rate Limits (429 RESOURCE_EXHAUSTED): Free-tier Gemini only allows 250K tokens/minute. During rapid testing, we burned through quotas instantly. We solved this by building a cascading model failover system and aggressively compressing input payloads by 80%.
  2. youtube-search-python + httpx Collision: The YouTube scraping library crashed on httpx >= 0.28 because of a deprecated proxies keyword. We surgically patched the library's internal HTTP layer to use the modern proxy syntax, then vendored the fixed version directly into our repo.
  3. Render 512MB RAM Ceiling: Local TensorFlow embedding models caused instant Out-Of-Memory crashes on Render's free tier. We completely pivoted to Google's cloud-hosted text-embedding-004 API, dropping RAM usage to near zero.

📚 What We Learned

  • How to engineer production-grade API rate-limit resilience using cascading model failover strategies.
  • How to construct and traverse Directed Acyclic Graphs for real-world prerequisite dependency resolution.
  • How to build Zero-Auth proxy scrapers that bypass restrictive API quotas using internal Protobuf endpoints.
  • How to patch and vendor third-party Python packages to resolve deep dependency conflicts in cloud deployment environments.
  • How to design glassmorphic, gamified UIs that keep learners engaged through visual progress tracking and micro-animations.

Built With

  • beautifulsoup4
  • clerk
  • docker
  • fastapi
  • google-gemini-2.5-flash
  • google-text-embedding-004-api
  • httpx
  • javascript
  • jspdf
  • networkx
  • postgresql
  • pydantic
  • pymupdf
  • python
  • react
  • react-flow
  • react-router
  • render
  • supabase
  • vercel
  • vite
  • youtube-search-python
Share this project:

Updates