No menus, no reports to dig through. Tell Jupeak what you want in plain Hindi or English — “who hasn’t paid me?”, “make a GST bill for Sharma”, “how’s business this month?” — and the AI does it, then shows you the numbers behind every answer.
Different jobs need different intelligence. We run ML.NET for tabular forecasting, anomaly detection on GL, embeddings for search, and Claude for natural-language reasoning on top.
SKU-level demand forecast, seasonal reorder points, cash-flow projection 90 days out. Trained per-tenant on your data, nightly.
Gradient-boosted model scores every lead 0-100 and every open deal by close-probability. Explainable — shows which features moved the score.
Watches GL postings, vendor rates, stock movements. Flags duplicate vouchers, vendor-rate spikes, stock-out risk, parties going past due.
Natural-language queries against your tenant. Asks for confirmation before any write operation. Drafts WhatsApp, emails, reports.
AI watches sale velocity, lead time, and supplier patterns. Creates purchase orders in draft, ready for your approval.
AI tracks purchase cadence, payment timing and WhatsApp response rate. Surfaces customers who are slipping before they leave.
Spots outlier GL entries, duplicate bills, vendor rates that spike outside normal bands. Before your auditor does.
AI writes personalised reminders in the customer's preferred language — Hindi, Tamil, Marathi — with their ledger snapshot + UPI QR pre-filled.
No black boxes. Here's the real stack behind Jupeak AI.
claude-sonnet-4-6 over Anthropic's commercial API, for the assistant, drafting and document reading. Runs per tenant with row-level security. Your data is not used to train any model. What is sent, and when →
Microsoft ML.NET for forecasting and scoring. Models trained nightly per-tenant. Runs on-prem friendly.
Hybrid embedding + BM25 for smart search across invoices, vouchers, tasks. Language-aware across English + 4 Indian languages.
Where your data goes, precisely: your books stay on our server. The forecasting and scoring models run there too and never leave it. When you use the assistant or document reading, that request is processed by Anthropic in the United States and deleted by them within 30 days — it is not used to train any model. Tenant isolation is enforced at the database row level and every AI action is in the audit log. Read the full detail, with sources →