SuperiorKe/order-to-cash-agent

Event-triggered order-to-cash agent for SME/Jua Kali manufacturers in Kenya — USSD/SMS/Voice via Africa's Talking, M-Pesa, and an owner-facing LiveKit voice assistant. # Order-to-Cash Agent An event-triggered agent for SME and Jua Kali manufacturers. It takes an order, confirms and schedules it, collects payment by M-Pesa, and chases late payment with an escalating SMS then a Voice call. Built for the Africa's Talking Manufacturing Hackathon, Nairobi, 27 Aug 2026. See `BUILD_PLAN.md` for the full design, the call sequences, and the demo script. See `AI_COMPONENTS.md` for the two AI systems in this build: the order-parsing brain and the owner voice assistant, "Friday". ## Architecture ```mermaid flowchart TD CPhone["Customer's phone USSD / SMS / Voice"] ATChannels["Africa's Talking (USSD + SMS + Voice, the spine)"] subgraph Server["Node server (src/)"] Intake["Intake routes/ussd.js, routes/sms.js"] Parse["Parse claude.js → DeepSeek V4 Flash (OpenRouter, tool use)"] Orders["orders.js / invoices.js"] Agent["Collections tick agent.js (cron, every minute)"] ATjs["africastalking.js (SMS + Voice out, retry/backoff)"] VoiceRoute["routes/voice.js (dial-plan XML)"] Mpesa["mpesa.js / routes/mpesa.js (OAuth, STK push, callback)"] Dashboard["routes/dashboard.js"] Api["routes/api.js"] end DB[("Postgres customers · products · orders invoices · messages")] Daraja["Safaricom Daraja STK push + callback"] Friday["Friday, voice-agent/ separate Python process LiveKit + Groq STT/LLM/TTS"] OwnerBrowser["Owner's browser"] OwnerPhone["Owner's phone (voice)"] CPhone --> ATChannels --> Intake --> Parse --> Orders --> DB Orders -- "confirmation SMS" --> ATjs --> ATChannels DB --> Agent Agent -- "reminder SMS #1 / #2, owner alert" --> ATjs Agent -- "voice escalation" --> VoiceRoute --> ATChannels --> CPhone Orders -- "STK push, on demand" --> Mpesa --> Daraja --> CPhone Daraja -- "callback" --> Mpesa --> DB DB --> Dashboard --> OwnerBrowser DB --> Api Api Friday OwnerPhone ``` The customer never sees a screen. Every order arrives by USSD, SMS, or Voice through Africa's Talking, gets parsed and priced, and lives in one Postgres schema that both …