Pulse · Internal AI Agent

Internal multi-platform AI agent. Business users ask natural-language questions in Discord and get live data from internal APIs via OpenAI tool-use.

TypeScriptNode.js 22+OpenAI Tool-use APIDiscord SDK

Highlights

→Natural-language queries mapped to internal APIs via typed tools
→Adapter pattern: Discord live, Telegram + WhatsApp planned
→Multi-system integration: balances across 4 external supplier APIs
→Slash commands as escape hatch when natural-language is overkill

Context

Business users (ops, marketing, support) needed live answers from internal data without filing a ticket or waiting for a dashboard build. Things like "tickets opened today", "top 5 products this week by revenue", "upsell performance last month".

Problem

A custom dashboard per question doesn't scale
A SQL prompt-to-output toy isn't safe. Needs typed access, not raw queries.
Users live in Discord, not in a custom UI
Same questions need to work later in Telegram and WhatsApp without rewriting the agent

What I built

A proper agent loop with typed tools, not a prompt-in / text-out wrapper.

Architecture

Agent loop in core/agent.ts: handles tool dispatch, conversation state, retries
Typed API client in core/api.ts: every call typed end-to-end, no raw HTTP
Tools per capability: each business question is a tool with a schema, parameters, and a typed return
Adapter layer: Discord live today, Telegram + WhatsApp adapters planned. Same agent core, swap the I/O.
Slash commands: structured escape hatch like /orderstats, /productstats, /combostats when natural-language is overkill
Multi-system integration: checks supplier balances across 4 external APIs from the same bot. Real multi-system integration inside one agent.

flowchart LR subgraph Adapters DC[Discord · live] TG[Telegram · planned] WA[WhatsApp · planned] end subgraph Core["Agent core"] AG[Agent loop<br/>core/agent.ts] LLM[OpenAI tool-use] AG <-->|"messages + tool calls"| LLM end subgraph Tools["Typed tools"] TB[Business queries] TS[Supplier balances] end DC --> AG TG -.-> AG WA -.-> AG SC[Slash commands<br/>/orderstats …] --> TB AG --> TB AG --> TS TB --> API[Typed API client<br/>core/api.ts] API --> INT[Internal API] TS --> SUP["4 supplier APIs"]

Every LLM action passes through a typed tool boundary. Audit, rate limits, and access control live there, so the model never touches raw HTTP or SQL.

Stack

TypeScript, Node 22+
OpenAI tool-use API (function calling)
Discord SDK
Cleanly separated layout: adapters / commands / core / tools / config with per-tool docs

Outcome

Business users now self-serve answers in Discord. The same agent core will ship to Telegram and WhatsApp without a rewrite. The "tools first, prompts second" architecture means we can audit and rate-limit at the tool boundary instead of trusting the LLM with raw access.

flowchart TB title["⚠ Before"] U1[Business user] --> T[Ticket] T --> D[Dev team] D --> DB[Custom dashboard] DB -.->|"days / weeks"| U1 style title fill:transparent,stroke:transparent

flowchart TB title["✓ After"] U2[Business user] -->|"natural language"| AG[Pulse agent] AG --> TT[Typed tools] TT --> API[Internal API] API -->|"seconds"| U2 style title fill:transparent,stroke:transparent