Agents In Action
Accounting Agent
Our own Profex Accounting System is built to be run by an AI agent: it exposes its entire back office through the Model Context Protocol (MCP), so an assistant like Claude can operate the books in plain language.
Challenge
Accounting software is powerful but slow to drive: every invoice, bill, payment, journal and reconciliation means clicking through forms and knowing where each feature lives. Teams without a dedicated bookkeeper lose hours to manual entry and month-end busywork, and the data stays locked behind the UI — out of reach of the AI assistants people now work with daily.
Solution
We made the platform MCP-native. An AI agent connects over the Model Context Protocol and drives the full finance workflow through 90+ tools: draft and approve invoices and quotations, record bills and payments, post and reverse journal entries, reconcile bank accounts, and generate reports. Seven built-in playbooks walk the agent through real jobs: month-end close, bank reconciliation, bill-payment runs, aged-receivables follow-up and quarter-end tax review. Every action is scoped to an API key, logged, and reversible. A companion Telegram bot lets staff capture invoices, expenses and receipts on the go.
Tenant Concierge
An AI assistant built into a tenant-management platform: landlords and butlers handle daily operations by chat on WhatsApp, Telegram or Feishu.
Challenge
Landlords managing multiple units juggle repetitive inquiries (room details, contract terms, vacancy status) across phone calls, messages and spreadsheets. New-lease onboarding means tedious manual data entry from signed contracts, and properties without smart meters need someone to record every utility reading by hand each month.
Solution
Built a LangGraph tool-calling agent into the tenant-management system, running through a pluggable platform layer on WhatsApp, Telegram or Feishu. Landlords and butlers ask about rooms, tenancies, vacancies, contracts and utility readings in plain language (English or Chinese) and the agent answers by querying the live system. For a new lease they send a photo of the signed contract and a vision-LLM extracts the tenant, rent, deposit and dates to verify before saving; for utilities they photograph the meter and the model reads the value, logs it and calculates usage — no manual typing.
Charger Diagnostics Agent
An AI diagnostics agent inside an EV-charging platform: operators ask what's wrong with any charger or session in plain English, and it investigates the live system to find out.
Challenge
An EV-charging network runs on distributed hardware from different vendors, each speaking the OCPP protocol a little differently. When a charge fails, ops staff had to dig through OCPP status codes, database records and log stores across several systems to work out why. It was slow work that needed deep platform knowledge, so problems sat unresolved.
Solution
Built an agentic investigator into the admin console. An operator describes a symptom and anchors it to a charger, session, transaction or user; the agent runs a diagnostic loop (a relevance gate, then playbook-guided reasoning over 12 tools that query the live database and the OCPP log store) and streams its progress live before producing a plain-English diagnosis and a handoff report. Seven root-cause playbooks (hardware fault, firmware, config, network, user error, software bug) steer it, all personal data is redacted before it reaches the model, and it's multi-LLM with side-by-side model comparison and per-investigation cost tracking.
