DSA-AI-Agent (Curator)

• 

  Website_UI

Inspiration

Museums. We built a Museum Sentinel: a calm, curator-style guide that teaches data structures and algorithms the way a docent leads a visit—structured, grounded in a fixed collection, and tuned to each visitor—not a generic web-wide chatbot.

What it does

• Answers from a read-only knowledge base (Backend/dsa_mock_data.json) so explanations match curated material.
• Personalizes using user_memory.json (skills, struggles, preferences, interests, progress).
• Reflects after each turn: updates the profile and appends sessions to history_log.json (summary, takeaways, mood).
• Loads the last three sessions into context so the assistant can connect threads.
• Keeps short-term dialogue with LangChain ConversationSummaryBufferMemory.
• Web stack: FastAPI backend + Next.js chat UI, with SSE streaming when available.

How we built it

Challenges we ran into

• “Failed to fetch” — The API was not running, NEXT_PUBLIC_API_URL pointed at the wrong base URL, or CORS rejected the request because the page was opened from the Network/LAN URL (e.g. http://10.x.x.x:3000) instead of http://localhost:3000 or http://127.0.0.1:3000.
• Windows PowerShell — && is often invalid for chaining commands; we used ; or ran commands on separate lines.
• Next.js dev — Only one next dev per project; a second instance fails until whatever owns port 3000 is stopped.
• Latency vs memory — Reflection (profile + history_log) means extra Gemini calls; we return the main answer first and run some persistence in background threads on the API path so the UI stays responsive.
• Quotas — Gemini rate / quota limits; we mitigated with ordered multi-key failover when keys belong to different Google Cloud / AI Studio projects.

Accomplishments that we're proud of

• Three-layer memory — Curated KB + mutable profile (user_memory.json) + session log (history_log.json) + ConversationSummaryBufferMemory, without ever writing to the read-only KB file.
• Streaming UX — POST /ask/stream (SSE) with a clean fallback to POST /ask when streaming is unavailable.
• Operable repo — GET /health, configurable CORS_ORIGINS, and a single root README.md covering setup, HTTP API, environment variables, and Windows troubleshooting.

What we learned

• Grounding an LLM on a local JSON corpus while updating learner state in separate files (clear separation of concerns).
• LangChain chat memory plus structured post-turn reflection wired through a real FastAPI → browser stack.
• How to tell Failed to fetch / CORS apart from HTTP 4xx/5xx and backend 500 responses when debugging from the frontend.

What's next for DSA-AI-Agent

• Chunked / indexed retrieval as dsa_mock_data.json (or the KB) grows.
• Multi-user isolation — auth or per-device user_memory.json / history_log.json.
• Light evaluation — a small golden set of DSA prompts for regression checks.
• Deployment — containers + hosted frontend; secrets only in environment, never in git.
• Pedagogy — guided exhibit-style paths, quizzes, or spaced repetition driven off the profile in user_memory.json.

Built With

• geminiapi
• langchain
• next.js
• pydantic
• python
• react
• typescript