InsightAI

Complaint intelligence platform: CSV in → ML classify → review queue → SLA analytics.

Stack: FastAPI · Streamlit · scikit-learn (tfidf-logreg-v2) · SQLAlchemy · SQLite
Version: 2.1.0

---

Live demo

Surface	URL
Dashboard	https://insightai-ui.onrender.com
API	https://insightai-api.onrender.com
API docs	https://insightai-api.onrender.com/docs
Health	https://insightai-api.onrender.com/health
Source	github.com/kajalmishra-dev/isi-ds-insightai

Free-tier hosts sleep after idle. First request after sleep can take ~30–60s.

Quick demo

1. Open the dashboard - wait for Online in the sidebar
2. Download sample CSV from the sidebar (or use data/sample_upload.csv)
3. Upload → watch the job finish
4. Check Overview, triage Review Queue, try Live Classification

---

What it does

• Async CSV ingestion with job progress, idempotent content-hash reuse, and retry
• TF-IDF + Logistic Regression classification with confidence → Needs Review
• Clear winners (top-1 vs top-2 margin) skip review even when max-prob is soft
• Ops dashboard: KPIs, category mix, SLA (% resolved in 24h), explorer + CSV export
• Human triage (Approve / Reject) with feedback counted for retraining
• Optional API-key auth, CORS, request IDs, /health + /ready

---

Architecture

Browser  →  Streamlit UI  →  FastAPI /api/v1
                                │
                   ┌────────────┼────────────┐
                   ▼            ▼            ▼
              ML (joblib)   SQLite DB   Background jobs

Hosted on Render (render.yaml): insightai-api + insightai-ui.

---

Local run

pip install -r requirements-dev.txt
# model artifacts ship in ml/artifacts/ - retrain only if needed:
#   python scripts/generate_training_data.py && python -m ml.train
uvicorn backend.main:app --reload
# other terminal
streamlit run frontend/app.py

Local	URL
Dashboard	http://127.0.0.1:8501
API docs	http://127.0.0.1:8000/docs

Docker:

docker compose up --build

Staging with auth:

$env:API_KEY="replace-me"
docker compose -f docker-compose.yml -f docker-compose.prod.yml up --build

Reset junk local data: python scripts/reset_local_db.py → restart → upload data/sample_upload.csv.

---

API (v1)

Method	Path	Description
GET	/health	Liveness
GET	/ready	DB + model ready
POST	/api/v1/upload	CSV → job (202, or 200 if duplicate content)
GET	/api/v1/jobs/{id}	Job status
GET	/api/v1/jobs	Recent jobs
POST	/api/v1/jobs/{id}/retry	Retry failed job
GET	/api/v1/analytics/summary	KPIs + insights
GET	/api/v1/complaints	Filter / search / paginate
GET	/api/v1/complaints/export.csv	Export
POST	/api/v1/predict	Classify one text
POST	/api/v1/complaints/{id}/review	Human triage

CSV columns: text, created_at, resolved_at

Auth (optional): AUTH_ENABLED=true + API_KEY → header X-API-Key.

---

Configuration

Copy .env.example → .env.

Variable	Purpose
AUTH_ENABLED / API_KEY	Protect /api/v1/*
REQUIRE_AUTH_IN_PRODUCTION	Force auth when ENVIRONMENT=production|staging
CONFIDENCE_THRESHOLD	Soft max-prob review cutoff (default 0.32)
CONFIDENCE_MARGIN	Clear winner margin to skip review (default 0.10)
API_BASE_URL	Frontend → API (Render UI uses the public API URL)

---

Deploy (Render)

Blueprint: render.yaml

1. Push main to GitHub
2. Render → New → Blueprint → this repo
3. Apply → wait for insightai-api + insightai-ui
4. Open the UI URL above

UI must set API_BASE_URL=https://insightai-api.onrender.com.
SQLite on free instances is ephemeral (redeploy clears demo data).

---

ML notes

• Synthetic training data (data/complaints.csv, 240 rows) - not customer data
• Demo upload (data/sample_upload.csv, 48 rows) is held out (no train overlap)
• Winner selected via holdout macro-F1 (ml/artifacts/experiments.json)
• Soft probabilities (~0.3–0.5) on a 4-class logreg are expected

pytest tests/ -v

---

Layout

backend/     FastAPI, auth, jobs, analytics
frontend/    Streamlit ops dashboard
ml/          Train / infer + committed artifacts
data/        Train + sample upload CSVs
docs/        Feature guide PDF
scripts/     Data + helper scripts
tests/       API / ML / product contracts

---

Limitations

• In-process jobs (not a durable worker queue)
• SQLite default (Postgres on the roadmap)
• Free Render sleep + cold start
• Metrics are demo-scale, not production customer performance

---

License

Private / portfolio project unless otherwise noted.