CareerSage

Inspiration

Across Nigeria and Africa, students often struggle to make informed career decisions due to limited counseling resources, inconsistent guidance, and lack of personalized advice. Many rely on generic quizzes that don't consider their real-life constraints — financial struggles, university strikes, power and internet issues, or the pressure of supporting family.

As a Nigerian student studying Cybersecurity at the University of Uyo, I've seen firsthand how difficult it is to access realistic, culturally-aware career planning tools. I built CareerSage to solve that problem: an AI-powered career counselor that truly understands the African student experience.

Key Features

Smart Advisor (Gemini 3.1 Pro Preview) Creates two personalized career roadmaps:
- Practical Pathway (6–18 months): quick wins & immediate stability
- Growth Pathway (3–5 years): long-term vision Includes salary charts and demand meters.
Voice Counseling (Gemini Live API) Natural voice conversations that answer real-time questions like JAMB dates, scholarship deadlines, and salary data — powered by gemini-2.5-flash-native-audio-preview-12-2025 with raw PCM audio streaming and Google Search grounding.
Career Simulator Experience a "day in the life" of any career using AI-generated workplace images and interactive roleplay with scene editing — powered by Gemini 3 Pro image generation.
Agentic Research Chat Fact-checks scholarships, tuition fees, deadlines, and job trends with live web search citations — powered by Gemini 3 Flash Preview with Google Search grounding.
Local Save System Saves sessions and roadmaps locally without storing data on external servers.

How I Built It

CareerSage uses a multi-model orchestration strategy across different modalities:

AI Models:

gemini-3-pro-preview (Gemini 3.1 Pro Preview) — deep reasoning, structured career roadmap generation, and the agentic research chat
gemini-2.5-flash-native-audio-preview-12-2025 — real-time voice counseling with PCM audio streaming
gemini-3-pro-image-preview — career vision image generation and scene editing for the Career Simulator
gemini-2.5-flash-image — simulation image editing

Frontend: React 18, TypeScript, Tailwind CSS Visualization: Recharts + Lucide icons Architecture: Modular multi-agent workflow using custom hooks and WebSocket-style voice streaming Data Handling: Client-side only (no backend database) for user privacy Testimonial API: Vercel serverless function writing to GitHub via Octokit, with rate limiting, input validation, and XSS sanitization CI/CD: GitHub Actions running TypeScript checks and deploying to Vercel on every push to main

The intake flow is a 7-step conversational form collecting name, current situation, interests, constraints, dreams, fears, and optional document uploads (WAEC results, JAMB score, CV). The strategy engine then runs a 3-stage pipeline: profile analysis → parallel pathway generation → result delivery.

Challenges I Ran Into

Designing advice that genuinely accounts for Africa's real constraints — ASUU strikes, NYSC, the HND vs BSc divide — not just generic career templates
Implementing interruption-aware voice sessions with raw PCM audio at 16kHz input / 24kHz output with scheduled audio buffering
Managing parallel AI calls (two pathways generated simultaneously) without blocking the UI
Ensuring CV and transcript privacy while still passing file data to the model as inline base64
Building a function-calling loop for the Career Simulator so the AI can both narrate and trigger image generation in the same turn

Accomplishments I'm Proud Of

Built a career counselor that doesn't give generic advice — it accounts for your real financial situation, family pressure, and local education system
Voice mode with interruption support and real-time search grounding — it can tell you current scholarship deadlines mid-conversation
The dual-pathway system gives students both an immediate option and a long-term vision, because life rarely fits one timeline
A complete serverless testimonial pipeline with proper security headers, CORS validation, and rate limiting

What I Learned

How to orchestrate multiple Gemini models into a single coherent user experience
How to design AI interactions that are culturally sensitive and locally grounded
Advanced React patterns for managing multimodal state across advisor, voice, simulator, and research modes
Raw PCM audio streaming and scheduling with the Web Audio API
Building for low-resource environments — slow networks, older Android devices, inconsistent power