IndustryF&B SaaS / Coffee chain / AI-Native

NiMCoffee

One cup of coffee, eight AI agents — ordering, supply chain, and operations all orchestrated by an Agent core.

About this page

The NiMCoffee system is fully built and has passed 30+ end-to-end tests. www.nimcoffee.com is currently being prepared for launch — contact us for a sandbox account or a private-deployment plan.

In short

NiMCoffee is an AI-native operating system for coffee shops, aimed at chain owners, store managers and front-line service staff. It treats the AI Agent as the business core rather than a bolt-on: any operation can be carried out in natural language, the agent orchestrates backend services through function calling, and the GUI only assists visualization. Eight Butler roles cover analytics, marketing, content, customer service and knowledge, running scheduled self-checks. Its most distinctive trait is that every pricing action passes through one shared `pricing_guard` validator, backed by a one-click kill switch.

What the agent takes off people's plates

Core capabilities

01

AI-Native core

The AI Agent is the business core, not an add-on. Function Calling + Agentic Workflow orchestrate every backend service; any operation can be completed by talking, with the GUI only assisting visualization.

02

Butler 8-role self-evolution

A Director coordinates Analytics / Marketing / Content / Customer-Service plus Conversation-Coach / Knowledge-Curator / Search-Optimizer. Daily 23:00 feedback mining, 03:30 hallucination scan, weekly 23:30 rule decay — all AI-driven.

03

Hybrid Fast-Path

Local product cache + slim tool set + keyword fast paths. 30/30 scenarios passing, average 3,068 tokens per turn, 92% token reduction vs. pure-LLM routing.

04

6 scenes · 33 tools

Scene-scoped tool sets: general / ordering / support / knowledge / management / channel-bot. Tools are dynamically pruned by role and scene — store managers see 33, channel bots read-only 15.

05

Multi-surface

Web (Nuxt SSR) + store mini-program (uni-app) + product mini-program + channel bots (WeChat OA / WeCom, AES-encrypted). The store mini-program has its own chat session and isolates manager vs. customer views.

06

Kill-switch & pricing guard

Base pricing is owner-only; AI marketing can move within ±20%, single coupon ≤¥50, daily budget ≤¥300, auto-refund ≤¥30 × 5/day. All pricing paths share a single `pricing_guard` validator (13 pytest tests). One-click kill-switch disables AI rules instantly.

Side by side

AI as the business core vs a traditional F&B SaaS plus manual work

DimensionTraditional F&B SaaS + manual workNiMCoffee (AI as the business core)
What sits at the centre of the systemGUI forms are the centre; AI is a chat plug-in added laterThe AI agent is the core; operations run through natural language and function calling, with the GUI only assisting visualization
Who does the daily operations workManagers count stock by hand, coupons go out on gut feel, service staff answer the same question over and overEight Butler roles run in parallel: daily 23:00 feedback mining, 03:30 hallucination scan, weekly 23:30 rule decay
How long a promotion takes to launchChange code, redeploy, canary — and wait for the engineering queueThe owner states the intent in plain language, AI Marketing drafts within the guardrails, then publish or kill-switch in one click
Controlling cost and riskPricing logic is scattered across entry points and reviewed by hand; once AI is bolted on, token cost grows linearly with conversation volumeEvery pricing path shares one `pricing_guard` validator with a kill-switch, and the hybrid Fast-Path architecture cuts per-turn token use by 92% versus pure-LLM routing

Quantified before / after

Measured impact

Owner / CEO
BeforeWait for the manager to send a monthly report — read it on the 5th
AfterAI daily report auto-generated 6:00 AM; ask anything from mobile any time
Store manager
BeforeManual stock count + Excel consumption tracking
AfterAI flags anomalies, deduplicates synonyms (milk / fresh milk / whole milk → fresh milk), warns 3 days before expiry
Marketing owner
BeforeIssue coupons by gut feel without knowing the actual lift
AfterAI Marketing builds coupons within ±20%, reviews same-day, failed rules auto-decay
Customer service
Before"What's available today?" "Can I have it iced?" — 100× a day
AfterChannel bot answers via 15 read-only tools, escalates the rest — humans focus on complaints
Engineering team
BeforeFestival promo push = code change + redeploy + canary
AfterOwner speaks the intent in natural language → AI Marketing proposes within guardrails → one-click publish / kill-switch

Subjects and capability matrix

Covered modules

Director

Coordinate / decide

Analytics

Trends / anomalies

Marketing

Pricing / coupons

Content

Posts / posters

Customer service

Tickets / refunds

Conversation Coach

Feedback → rules

Knowledge Curator

Knowledge gap scan

Search Optimizer

Synonyms / intent

Verifiable facts

Numbers and sources

30 of 30 end-to-end scenarios pass, averaging 3,068 tokens per turn — a 92% reduction versus pure-LLM routing.

Self-reportedMeasured on NiMCoffee's own 30-scenario internal test set, baselined against pure-LLM routing; not independently verified2026-08-01

The tool registry spans 6 scenes and 33 tools, pruned dynamically by role and scene: store managers see all 33, channel bots see 15 read-only.

Self-reportedNiMCoffee system configuration notes, verified 2026-082026-08-01

The Butler team consists of 8 roles with three fixed self-check jobs: daily 23:00 feedback mining, 03:30 hallucination scan and weekly 23:30 rule decay.

Self-reportedNiMCoffee system configuration notes, verified 2026-082026-08-01

Pricing guardrails allow AI marketing ±20% movement, a ≤¥50 single coupon, a ≤¥300 daily budget and auto-refunds of ≤¥30 up to 5 per day; all pricing paths share one `pricing_guard` validator covered by 13 pytest cases.

Self-reportedNiMCoffee system configuration notes, verified 2026-082026-08-01

Where it fits

When this approach does not apply

  • The 30/30 scenario pass rate and the 3,068-token average come from YGG's own internal test set, not from long-run trading data in a real shop. Real customers phrase things in a far longer tail, so both hit rate and token consumption will be worse than in the test environment.
  • The agent's autonomy is deliberately fenced in: base pricing is owner-only, and AI marketing may act only within ±20%, a ≤¥50 single coupon and a ≤¥300 daily budget. Scenarios that expect the AI to price freely or commit large budgets on its own are out of scope.
  • Multi-surface access depends on the WeChat ecosystem: the WeChat OA and WeCom channel bots require the merchant's own verified accounts and AES configuration. Without them that capability is unavailable, while the other surfaces keep working.
  • A single shop with low daily order volume gains little. Self-evolving roles, knowledge-gap scanning and rule decay all feed on accumulated conversation volume; with too few samples a manager's own judgement is usually the cheaper answer.
Written by: YGG Technology solutions teamPublished: 2026-05-08Last reviewed: 2026-08-01