Private analyst for mid-size Malaysian e-commerce sellers
Challenge: Mid-size Malaysian e-commerce sellers (solo or 2–5 person teams, RM30K–RM300K/month GMV) are the busiest people in their business — packing, customer service, ads, fulfillment, all at once. Freelance e-commerce consultants cost RM3K–10K/project, raise confidentiality concerns (often serving competitors), and speak in jargon. Platform-native dashboards show revenue but not net margin after discounts, returns, and shipping. The seller watches sales grow while profit silently erodes.
Solution: Ecommerce Analyst is a private analyst-in-a-box that lives inside the seller's own instance. It answers plain-language questions about their transaction data and delivers a proactive Monday morning brief — without sharing data with any human consultant, at a cost below any part-time hire.
Free-form chat grounded in the seller's transaction CSV. Questions like "what's eating my margin?", "is my 15% Raya discount profitable?", "which state is losing me money?" — answered in plain language with specific numbers.
Auto-generated every Monday 7am: 3 wins, 3 worries, and 1 decision to make this week. The seller reads it over coffee; clickable findings open the chat to drill deeper.
Preset buttons for the most common seller questions: profit by category, state shipping economics, returns by SKU, discount ROI, top customer cohort, delivery performance. Zero typing needed.
Drop the seller's platform transaction export. App auto-detects columns, remaps regions to Malaysian states, drops platform-only columns the seller has no visibility into.
Every Q&A and Monday brief is saved and searchable. Export as PDF to share with accountants, partners, or family members running the business.
Two-panel design optimized for the seller's phone. Works between packing runs at 11pm without a desktop, without a data analyst, without a consultant.
The live version uses Gemini 2.5 Flash to convert plain-language seller questions into pandas queries against the CSV, then translates the numeric result back into a one-paragraph answer in Malay-flavored English.
Every Monday 7am, the scheduler pulls the last 7 days of data and the LLM writes a scannable brief — 3 wins, 3 worries, 1 decision. Each finding is grounded in a specific number from the data.
The system prompt enforces confidentiality: no platform names mentioned, no data leaves the seller's instance, no cross-client pattern references. The analyst serves only the seller it's deployed for.
This portfolio demo pre-computes 6 quick-action answers + one Monday brief + 5 example asks from the full 34,500-row dataset. No live LLM call needed — demonstrates the output shape without any API cost.
data_prep.py cleans the transaction CSV — remaps region to 16 Malaysian states with realistic distribution, regenerates delivery times with Malaysian logistics spread (1 day Klang Valley → 7 days East Malaysia), drops platform-only columnsprecompute_answers.py computes all 6 quick-action answers + Monday brief + example asks from the full 34,500-row dataset, outputs single answers.jsonanswers.json is inlined directly into demo.html — no CORS issues, no external fetch, works on file:// and GitHub Pages identicallypython data_prep.py — Clean and localize the transaction CSVpython precompute_answers.py — Generate data/answers.json with all demo responsesopen demo.html — Launch the self-contained static demo in a browser