Inspiration Startup founders rarely need another chatbot that gives advice, they need operations to actually get done: find vendors, track inventory, know when stock is low, pay suppliers, and notify them. all without juggling five apps. We saw founders search suppliers on Instagram/Facebook marketplaces, track stock in spreadsheets, and only realize they were out of beans when the cashier kept selling. Generic AI answers “you should reorder” but does not connect marketplace discovery → stock levels → budget check → DOKU payment → vendor message.

FoundrOS COOPilot AI was built to close that loop, an autonomous multi-agent operations employee on Telegram, with persistent business memory, real payments, and trust boundaries (only registered suppliers get paid).

What it does

FoundrOS COOPilot AI is autonomous operational intelligence for small businesses. It goes beyond recommendations and executes workflows.

  1. Company onboarding & memory Detailed company profile (legal entity, brand, industry, address, budget, goals, GPS). Stored in Mem9, scoped per user, so agents remember your business.
  2. Vendor discovery (two channels) Users can find suppliers through: Channel How it works Map-based (OSM) /cari_supplier — top nearby suppliers with address & distance from office GPS Social media marketplace (Repliz, In this Product we not implement it because the limitation of time ) Search/listing-oriented vendor options from social commerce marketplaces connected via Repliz API—where many UMKM actually source packaging, beans, and supplies today Discovered vendors can be saved to the supplier registry; only registered vendors enter payment flows.

  3. Stock monitoring & POS integration FoundrOS COOPilot tracks inventory per SKU/product:

When the cashier records units sold (POS / kasir input), COOPilot automatically deducts stock Current level is compared to a defined capacity or baseline Low-stock rule: when stock falls below 20% of capacity: $$current_stock / capacity < 0.20$$ COOPilot sends a low-stock alert and routes the founder straight into supplier payment. linking the right registered vendor, budget check, DOKU invoice, and vendor notification (Telegram / WhatsApp draft).

  1. Multi-agent payment & communication For /bayar or triggered reorder flows:

Strategy → Memory → Finance → Payment (DOKU) → Communication → Operations log

  • Finance verifies $$payment_amount ≤ budget_available$$
  • Payment generates a real DOKU checkout link
  • Communication drafts can send Telegram invoices to vendor IDs
  • Social agent (Repliz) supports marketplace/social workflows where configured
  • Planning without payment /rencana runs strategy and operational planning without opening DOKU—useful for growth goals separate from procurement.

How we built it

Stack: Python, Telegram Bot, FastAPI, Sumopod (LLM), Mem9 (memory), DOKU (payments), Repliz (social/marketplace), OpenStreetMap (geo discovery).

Telegram / POS (sales events) ↓ Intent router + Orchestrator ↓ ┌───────────────────────────────────────────────────────┐ │ Strategy │ Memory │ Inventory │ Social (Repliz) │ │ Finance │ Payment (DOKU) │ Communication │ Ops log │ └───────────────────────────────────────────────────────┘ ↓ Mem9 (profile, suppliers, stock history, execution logs)

Design choices :

  • Supplier registry = payment whitelist (anti-hallucination)
  • Dual discovery: factual geo (OSM) + social marketplace (Repliz)
  • Inventory as a first-class signal: cashier sales → stock delta → threshold agent → payment orchestration
  • Sequential operational feed on Telegram so judges and founders see autonomy step-by-step ## Challenges we ran into
  • Payment hallucination — global memory once suggested wrong vendors; fixed with scoped Mem9 tags and registry-only payees.
  • Stock sync reliability — POS events must be idempotent and fast; we designed clear “units sold” input and atomic stock updates before triggering alerts.
  • Threshold false alarms — capacity must be set per SKU; we use explicit < 20 % <20% rule and tie alerts to registered suppliers only.
  • Telegram UX — typos (/carisupplier), GPS skips, and 409 multi-instance bots required aliases, /lokasi, and single-bot discipline.
  • API instability — Sumopod model balances, Overpass 406 responses, DOKU HMAC signing—each needed fallbacks and tests.

Accomplishments that we're proud of

  • End-to-end execute, not advise: Mem9 + DOKU + vendor messaging in one product story
  • Dual vendor discovery: maps and social marketplace (Repliz)
  • Inventory-aware autonomy: cashier sales → auto stock deduction → <20% alert → guided supplier payment
  • Multi-agent feed that makes autonomy visible and trustworthy
  • Payment gating and scoped memory after real user feedback on wrong vendor names
  • Integration tests for Mem9, DOKU, and gated payment workflows ##What we learned
  • Operations products win when signals (stock, location, budget) connect to actions (pay, notify, log)—not when chat is isolated
  • Marketplace reality: many vendors live on social commerce; Repliz integration matches how UMKM actually buy
  • Inventory thresholds are a natural trigger for autonomous procurement—if alerts do not link to payment, founders ignore them
  • Multi-agent splits improve debugging, demos, and safety (Finance before Payment)
  • Autonomy without whitelists is dangerous; registry + profile completeness are non-negotiable ## What's next for FoundrOS COOPilot AI
  • Deeper Repliz marketplace search — richer filters (category, region, rating) and one-tap “save as supplier”
  • Full POS connectors — beyond manual cashier input: webhook integrations with popular POS/cashier apps
  • Predictive replenishment — forecast depletion date from sales velocity before hitting 20%
  • DOKU webhooks — auto-confirm payment and notify vendor without manual follow-up Streamlit / web ops dashboard — stock board, feed, memory, finance in one view
  • Production deploy — VPS, Telegram webhooks, environment-separated Mem9
  • WhatsApp Business API — outbound delivery for invoice messages, not drafts only

Built With

  • cursor
  • doku
  • fastapi
  • map
  • mem9
  • python
  • repliz
  • sumopod
Share this project:

Updates