Doctor Agent - File medicine specialist

Inspiration

I’ve always seen colleagues “solve” messy storage by resetting their computers because it’s faster than cleaning up: downloads are chaotic, installers pile up, duplicate assets spread everywhere, and nobody wants to risk deleting something important. The real pain isn’t disk space it’s decision fatigue + fear of breaking things**. Doctor Agent was inspired by the idea that file cleanup should feel like a safe medical checkup

What it does

Doctor Agent is a local file organizer that helps you safely clean and reorganize your machine:

• Multi-folder scan: pick multiple folders and list files with metadata (size, modified date, paths).

• Diagnosis with local AI (LM Studio): classifies selected files and returns:

  • category, confidence, and a short reason
  • a suggested action: Keep, Quarantine, or Cloud (Google Drive)

• Safe treatment:

  • Quarantine moves files into a Quarantine/ folder inside each selected root.
  • Cloud uploads to Google Drive and then quarantines the local copy.

• Review-first UX: the user stays in control and can apply actions file-by-file or in batch.

How I built it

• Frontend: React + TypeScript

  • Uses the File System Access API to pick folders and move files locally.
  • Shows a “medical” workflow UI: Intake → Diagnose → Apply treatment → Report.
  • Image preview for supported image files.

• Backend: Fastify + TypeScript

  • /classify builds a strict prompt and calls LM Studio’s OpenAI-compatible API.
  • /drive/upload uploads files to a Google Drive folder (service account), then the frontend quarantines locally.

• AI: LM Studio local server

  • Runs a local model so classification doesn’t require cloud inference.

• Deployment approach

  • Frontend is hosted (Vercel).
  • Backend runs locally and is exposed through a VS Code forwarded port, so the hosted UI can talk to the local machine.

Challenges I ran into

• Model output reliability: local LLMs sometimes return markdown fences, partial JSON, or duplicate outputs. We had to harden the prompt and implement robust JSON extraction/parsing.
• Latency and UX: local inference can be slow, especially on first load. We had to design the experience so “waiting” feels expected and safe.
• Cloud upload correctness: handling file names, MIME types, and ensuring the file actually lands in the right Drive folder.
• Keeping actions safe: moving files is inherently risky; the quarantine design and “review-first” approach were essential.
• Deployed UI + local services: bridging a hosted frontend to a local backend without ngrok required careful coordination (CORS, forwarded URLs, environment config).

Accomplishments that im proud of

• Built a real scan → diagnose → treat pipeline that works end-to-end.
• Integrated local AI (LM Studio) instead of relying on external APIs.
• Implemented a Cloud + Quarantine workflow that is cautious by default (backup-first behavior).
• Made multi-folder scanning practical while keeping the UX clean and readable.
• Produced actionable, explainable output (category + confidence + reason), not just labels.

What I learned

• The hard part of AI file cleanup isn’t classification—it’s trust and safety.
• Strict prompting isn’t enough; you need defensive parsing and careful failure modes.

What's next for Doctor Agent - File medicine specialist

• More reliable structured output: function-calling style schemas or stronger validation + retries.
• Better treatment plan: “prescriptions” per category (e.g., duplicates, large installers, old downloads).
• Search + filters: sort by size, age, extension, category, and risk level.
• Deduplication support: detect duplicates via hashing and suggest safe consolidation.
• Safer rollback: an “undo” log and restoration from Quarantine.
• Background operations + progress UI: better loading states and operation-level progress (uploading, diagnosing, moving).
• Optional cloud providers: Drive today, but expand to Dropbox/OneDrive later.

Built With

• javascript
• react
• typescript