Skip to main content

Technical Architecture

Last Updated: 2026-06-04

Core Innovation​

Not just ChatGPT wrappers - we build real adaptive learning systems

1. Real-Time AI Question Generation​

  • Generate infinite personalized questions on-demand
  • Contextual to learner's weak areas
  • Adaptive difficulty based on performance
  • Multi-modal (code, SQL, system design, MCQ, essays)

2. Algorithmic Adaptivity (IRT/BKT)​

  • IRT (Item Response Theory): Match learner ability to question difficulty
  • BKT (Bayesian Knowledge Tracing): Track probability of concept mastery
  • Real-time difficulty adjustment (not static question banks)
  • Mastery-based progression (can't advance until 80% correct)

3. Multi-Agent Architecture​

  • 10+ specialized AI agents (not one general chatbot)
  • Each optimized for specific task
  • Coordinated via central orchestrator
  • See agent details below

AI Agent System​

Agent 1: Tutor Agent​

  • Role: Explain concepts, answer questions, provide hints
  • Model: Claude 3.5 Sonnet (now), fine-tuned Llama 3 (Month 6+)
  • Cost: ₹2-4 per interaction
  • Features: Socratic method, never gives direct answers, guides discovery

Agent 2: Mock Interview Agent​

  • Role: Conduct technical + behavioral interviews
  • Model: GPT-4 for conversational (voice-enabled)
  • Cost: ₹20-40 per session
  • Features: Real-time feedback, tracks body language (future), rates answers

Agent 3: Mentor Agent​

  • Role: Career advice, skill gap analysis, goal setting
  • Model: Claude 3.5 Sonnet + RAG (job market data)
  • Cost: ₹10-20 per session
  • Features: Personalized career paths, salary predictions, job recs

Agent 4: Code Review Agent​

  • Role: Review code quality, suggest improvements, best practices
  • Model: Fine-tuned CodeLlama
  • Cost: ₹5-10 per review
  • Features: Performance analysis, security checks, refactoring suggestions

Agent 5: Resume Agent​

  • Role: Build, optimize, ATS-check resumes
  • Model: Claude 3.5 Sonnet
  • Cost: ₹20 per review/build
  • Features: ATS optimization, keyword analysis, format templates

Agent 6: Project Guide Agent​

  • Role: Suggest projects, guide implementation, review outcomes
  • Model: GPT-4 + code execution sandbox
  • Cost: ₹10-30 per project
  • Features: Step-by-step guidance, code hints, debugging help

Agent 7: Assessment Generator Agent​

  • Role: Generate custom assessments from JD/topics
  • Model: Claude 3.5 Sonnet
  • Cost: ₹3-8 per assessment
  • Features: JD parsing, skill extraction, difficulty calibration

Agent 8: Feedback Agent​

  • Role: Detailed explanations on wrong answers
  • Model: Fine-tuned Llama 3
  • Cost: ₹4-5 per feedback
  • Features: Error pattern detection, concept gaps, practice recommendations

Agent 9: Strength/Weakness Profiler Agent​

  • Role: Continuous skill assessment, visual dashboards
  • Model: Python analytics + ML clustering
  • Cost: ₹1-2 per update (compute only)
  • Features: Skill heatmaps, growth tracking, peer comparisons

Agent 10: Hiring Match Agent​

  • Role: Match candidates to jobs based on skills
  • Model: Embedding-based similarity (Vector DB)
  • Cost: ₹2-5 per match
  • Features: Job scraping, skill mapping, application automation

Technology Stack​

Frontend​

  • Framework: Next.js 14 (React 18)
  • UI: Tailwind CSS + shadcn/ui
  • State: Zustand + React Query
  • Real-time: WebSockets (Socket.io)

Backend​

  • API: Node.js + Express (REST + GraphQL)
  • Auth: Clerk / Auth0 (SSO support)
  • Payments: Razorpay (India), Stripe (international)
  • Jobs: BullMQ (Redis-based)

Database​

  • Primary: PostgreSQL (user data, transactions)
  • Cache: Redis (sessions, leaderboards)
  • Vector: Pinecone / Weaviate (semantic search, job matching)
  • Analytics: ClickHouse (events, usage tracking)

AI/ML​

  • LLM APIs: Claude 3.5 Sonnet, GPT-4 (Phase 1)
  • Self-hosted: Llama 3 fine-tuned (Phase 2+)
  • Code Execution: Judge0 / Piston (sandboxed)
  • Embeddings: OpenAI text-embedding-3

Infrastructure​

  • Hosting: AWS / GCP (start with AWS)
  • CDN: Cloudflare
  • Monitoring: Datadog / New Relic
  • Logging: ELK Stack

AI Cost Optimization Roadmap​

Phase 1 (Month 1-6): API-Based​

  • Use Claude 3.5 Sonnet API
  • Cost: ₹1.50-2/question
  • Quick to market, no ML ops

Phase 2 (Month 6-12): Fine-Tuning​

  • Fine-tune Llama 3 on expert questions
  • Self-host on AWS EC2/GCP
  • Cost: ₹0.60/question (60% reduction)
  • Requires ML engineer

Phase 3 (Month 12-18): Custom Models​

  • Build smaller task-specific models
  • Distill from Llama 3
  • Cost: ₹0.30/question (80% reduction)
  • Requires ML team (2-3 people)

Phase 4 (Month 18+): Edge Optimization​

  • Cache common questions/feedback
  • Pre-generate question banks (indexed)
  • Cost: ₹0.10-0.20/question (90% reduction)

Goal: ₹24 crore AI costs → ₹6 crore (18% → 5% of revenue)

Open Source Strategy​

Year 1: Closed (Build & Validate)​

  • Focus on product-market fit
  • Validate algorithms work
  • Build community trust

Year 2: Open Algorithms​

  • IRT/BKT implementations (MIT license)
  • Question generation prompts
  • Dataset of 10K+ questions (Creative Commons)

Year 3: Selective Platform​

  • Frontend components (UI library)
  • API SDKs (Python, JavaScript)
  • Self-hosting guides
  • Keep core platform proprietary (non-profit IP)

Why selective:

  • Prevent for-profit clones
  • Maintain mission control
  • Balance transparency with sustainability

Security & Privacy​

Data Protection:

  • End-to-end encryption (user data at rest)
  • GDPR/CCPA compliant
  • Regular security audits
  • Bug bounty program (Year 2)

Code Security:

  • Sandboxed code execution (Judge0/Piston)
  • Rate limiting (prevent abuse)
  • DDoS protection (Cloudflare)

Privacy:

  • Minimal data collection
  • User controls (delete data anytime)
  • No selling data (non-profit = no incentive)

Scalability Plan​

100K Users (Month 12)​

  • 2-3 backend instances
  • PostgreSQL primary + 1 read replica
  • Redis single instance
  • Cost: ₹2-3L/month infra

1M Users (Month 18)​

  • 10+ backend instances (auto-scaling)
  • PostgreSQL sharding
  • Redis cluster
  • CDN for static assets
  • Cost: ₹10-15L/month infra

5M Users (Month 24)​

  • 50+ backend instances
  • Multi-region deployment (AWS Mumbai + Singapore)
  • Dedicated ML inference servers
  • Cost: ₹40-50L/month infra

Cost per user drops as we scale (economies of scale)