ThermIQ — Industrial Knowledge Intelligence
AI-powered knowledge platform for thermal power plants. Quantifies knowledge
gaps as ₹ crore operational risk.
Live: https://therm-iq.vercel.app (mirror: https://ghostunamused.github.io/thermIQ)
21-second intro (media/thermiq-intro.mp4), made with the /brag Claude Code plugin.
Architecture
- Frontend: static single-page app in
docs/ (served by Vercel and GitHub Pages)
- Backend: Vercel Functions (
api/*.js)
- Vector DB: Qdrant Cloud (collection
thermiq_chunks)
- Embeddings: Jina AI (jina-embeddings-v3)
- Generation: Gemini 2.5 Flash, with NVIDIA NIM and OpenRouter fallbacks
- Structured data: Firebase Firestore
- Knowledge graph: Neo4j Aura
- Pipelines: GitHub Actions (daily CEA outage fetch, gap scans, Drive ingest, Neo4j keep-alive)
Local setup (fresh clone)
git clone https://github.com/GhostUnamused/thermIQ.git
cd thermIQ
cp .env.example .env # fill in keys (same values as the Vercel dashboard)
npm install # deps for api/*.js
python -m venv .venv && source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r scripts/requirements.txt
Preview the frontend: python -m http.server 5173 --directory docs (it calls the
live Vercel backend). To run the API locally too: npx vercel dev (after npx vercel link).
Deploys are automatic: pushing to main redeploys Vercel (frontend + API) and GitHub Pages.
GitHub Actions secrets must mirror the keys listed in .env.example.
Claude Code plugins
The repo pins the brag and frontend-design plugins in .claude/settings.json. To install them for your
user account on a new machine: powershell -ExecutionPolicy Bypass -File scripts\install_claude_plugins.ps1
(Windows) or sh scripts/install_claude_plugins.sh (macOS/Linux).
Common tasks
Ingest benchmark documents (CEA standards — the yardstick):
python scripts/ingest_documents.py <pdf_path> guideline "CEA Tech Spec 500MW" <url>
For scanned/image PDFs use scripts/ingest_ocr.py instead.
Compute knowledge gaps (single source of truth for risk_scores):
python scripts/detect_gaps.py [--client NAME]
Fetch CEA outages manually: python scripts/fetch_cea_outage.py
Risk formula
risk_score_cr = criticality_score (1-5) × consequence_cr (₹ Cr) × exposure_score (0-1)
- criticality: 1-5, sourced to CEA outage frequency + CERC regulations
- consequence_cr: avg revenue impact from CEA outage records (₹5.0/kWh, LBNL/Ember 2024)
- exposure: 1 − best client-corpus match score
