📡 PULSE — Customer Intelligence Platform

Live customer & competitor intelligence for consumer brands, built at near-zero cost.

🔗 Live demo: sagnikpaulfirstclub.streamlit.app — 9,000+ real Play Store reviews of FirstClub, Zepto, Blinkit, BigBasket & Swiggy Instamart, refreshed daily. (The Ask-the-Data chat page is passcode-gated to limit API spend — code shared on request.)

What it does

Layer 1 — real public data. Scrapes Google Play reviews for a company and its competitors daily, tags complaint themes, benchmarks sentiment, detects emerging spikes, writes a Monday CEO digest with Claude, and answers free-form questions over the corpus.

Layer 2 — internal-data capabilities on clearly-badged synthetic data: RFM/KMeans segmentation, leakage-safe churn prediction (AUC 0.93), item-item recommendations, and AI-written per-segment win-back messaging.

Sample findings (July 2026 data)

• FirstClub has the lowest negative-review share of all five apps — 16.2% vs Zepto's 41.4% — and delivery-speed complaints are just 5.6% of its negatives vs 10–16% for competitors: the quality-first positioning is measurably landing.
• Refunds + support are quick-commerce's open wound — ~32% of negative reviews industry-wide, the #1 complaint macro-theme. FirstClub is best-in-class at 17.8%.
• The engine also finds the uncomfortable stuff: FirstClub's own #1 complaint theme is app UX/stability at 27.5% of negatives (~2.5× category norm) — onboarding and cart-sync specifically.

Engineering choices

Cost: ₹0 infrastructure (Streamlit Community Cloud + GitHub Actions free tier) · ≈ $1–2/month of Claude API. Sentiment = star ratings (the customer self-labels), theme tagging = keyword taxonomy first (~70–85% coverage), scikit-learn for all ML — Claude only touches what an LLM is irreplaceable for: labeling ambiguous reviews (Haiku, batched 50/call), the weekly digest, and the chat layer (Sonnet).

SQLite-in-repo instead of a hosted DB (single daily batch writer, read-only dashboard, data history versioned in git). GitHub Actions instead of an orchestrator (linear batch pipeline living next to its code). Synthetic-but-honest demo data instead of fabricated "real-looking" datasets.

Run it for YOUR company in 10 minutes

1. Edit config.yaml — company name, brand color, Play Store ID, competitor IDs. That's the whole generalization layer.
2. pip install -r requirements.txt
3. python scraper/play_store.py && python nlp/run_pipeline.py
4. python models/synth_data.py && python models/segments.py && python models/churn.py && python models/recommender.py
5. streamlit run app/Home.py

Deploy (free)

Push to GitHub → share.streamlit.io → New app → entry point app/Home.py → add ANTHROPIC_API_KEY and CHAT_PASSCODE in Secrets. GitHub Actions handle the daily scrape and weekly digest (add ANTHROPIC_API_KEY under repo Settings → Secrets → Actions, then run the one-click bootstrap-pipeline workflow once).

Structure

config.yaml          # the generalization layer — the only file that changes per company
scraper/             # Google Play review scraper -> SQLite
nlp/                 # taxonomy tagging, Claude batch labeling, aggregates, spike detection
models/              # synthetic data, segments, churn, recommender (demo layer)
llm/                 # CEO digest, ask-the-data chat, win-back copy
app/                 # Streamlit dashboard (7 pages)
.github/workflows/   # one-click bootstrap + daily scrape + weekly digest crons
data/                # SQLite DB + parquet outputs + digest archive (committed by CI)

---

Independent project on public data. Not affiliated with FirstClub or any company shown. Built by Sagnik Paul · BITS Pilani Hyderabad · paul1sagnik@gmail.com