shift-sage

The retiring expert that never clocks out: instant, cited fixes from your plant's own maintenance history.

Comment

Shift Sage

The retiring expert that never clocks out. Instant, cited fixes from your plant's own maintenance history.

We built Shift Sage for the $1,000 Industrial AI Hackathon. It targets a problem every small-to-mid manufacturer knows: when the night-shift tech hits an unfamiliar fault, the fix that saves four hours of downtime lives in someone's head, not in any searchable system.

Shift Sage architecture diagram

Quick demo

  1. Open the app and click Get started
  2. Click Load demo fault or paste: packing line 3 dead, swapped prox sensor and relay, PLC looks fine, still nothing
  3. Hit Diagnose and watch the agent trace search plant history
  4. See the ranked plan - step 1 cites Frank Martinez's undocumented east guard panel bypass from WO-2023-0847
  5. Click Generate work order and handover to close the loop
  6. Open the Dashboard to see Frank flagged at 95% retirement risk with 4 exclusive PL3 workarounds

No API keys required for the demo path. Add Gemini, Groq, or Mistral keys in Settings for live AI.

Setup

Local development

npm install
npm run dev

Open http://localhost:3000.

Docker

docker build -t app .
docker run -p 3000:3000 app

Open http://localhost:3000.

API keys (optional)

Visit /settings to paste free-tier keys:

Provider Use Sign up
Google Gemini Primary diagnosis over full asset history aistudio.google.com/apikey
Groq Fast streaming reasoning trace console.groq.com/keys
Mistral Failover for log cleaning console.mistral.ai/api-keys

Keys are stored in localStorage under shiftsage_api_keys and sent as request headers only.

Tech stack

  • Frontend: Next.js 14, TypeScript, Tailwind CSS, shadcn/ui-style components
  • Data: SQLite (better-sqlite3) with 18 seeded work orders across 8 assets
  • Retrieval: Keyword search plus asset alias matching (no heavy vector DB)
  • AI: Gemini 2.5 Flash (primary), Groq Llama 3.3 70B (streaming trace), Mistral (failover)
  • Deploy: Docker standalone output on port 3000

Architecture

Technician fault input
        |
        v
  Asset detection + keyword retrieval (SQLite)
        |
        v
  Live reasoning trace (SSE stream)
        |
        v
  LLM diagnosis (grounded to retrieved WOs only)
        |
        +---> Cited repair plan + workaround finder
        +---> Auto work order draft
        +---> Shift handover summary
        |
        v
  Manager dashboard (recurring failures, retirement risk)

Every recommendation must cite a work order ID from retrieved history. If history is too thin, Shift Sage refuses to guess.

Features

  • Instant cited diagnosis - ranked repair steps linked to specific past work orders
  • Hidden workaround finder - surfaces tribal knowledge from technician notes
  • Live agent reasoning trace - visible search, match, and rule-out steps
  • Auto work order and shift handover - cleans noisy input into CMMS-ready output
  • Voice fault capture - hands-free input via browser speech recognition
  • Manager knowledge-risk dashboard - recurring failures and retirement concentration

Participant

Aryan Choudhary - aryancta@gmail.com

Credits

  • Synthetic dataset inspired by MaintNet and Atlas CMMS log schemas
  • Research references: Dovient tribal knowledge report, Oxmaint CMMS guide, arXiv 2511.05311 (log cleaning agents), MDPI 2025 multi-agent PdM framework

Built With

Share this project:

Updates