MangaAssist Architecture DeepDive

Full HLD + LLD for the MangaAssist case study. Walks the path from anime-finisher → manga reading-order recommendation, through the orchestrator/RAG/tool-use stack, all the way to the cost envelope and scale-out plan.

Interview talking points

• Walk me through the architecture. Use these notes as your spine — they trace user → API gateway → orchestrator → tools/retrieval → LLM → trace store, with tradeoffs called out at each hop.
• Why SSE over WebSocket for chat? The transport-choice notes here cite latency, CDN compatibility, and reconnect semantics.
• One brain, many surfaces. The orchestrator-as-seam pattern is the load-bearing design choice; rehearse why it matters when you swap Streamlit → React without rewriting the brain.
• Where do you store traces? SQLite for the prototype, swap-pattern to DynamoDB / Postgres outlined here.
• Cost envelope. Per-request token math + cache-hit math sit in this folder; rehearse the back-of-envelope before the Cost Optimization section.

Files in this folder

Back to the home page.