Kaveen Kumarasinghe - founder of GPT Discord - LinkedIn

Enables real-time LLM interactions directly within Discord servers through a bot that parses user messages, routes them to language model backends (likely OpenAI GPT), and streams responses back into Discord channels with native formatting and threading support. Uses Discord.py or similar bot framework to hook into Discord's gateway API for message events, maintains connection pooling to LLM providers, and handles rate limiting across both Discord API and LLM service tiers.

Maintains conversation state across multiple Discord messages by fetching and indexing prior message history from channels, building a sliding-window context buffer that feeds into LLM prompts to enable coherent multi-turn interactions. Implements message deduplication, timestamp-based ordering, and optional summarization of older messages to stay within LLM context windows (typically 4K-128K tokens depending on model). Uses Discord's message fetch API to retrieve historical context and implements local caching to reduce API calls.

Intercepts Discord messages and classifies them as commands (e.g., !ask, /gpt) versus natural conversation, routing commands to specific handlers (summarize, translate, code-review) while passing natural messages to the LLM. Implements a command registry pattern where handlers are registered with argument schemas, validation rules, and permission checks. Uses regex or Discord's native slash-command API for parsing, with fallback to prefix-based commands for backward compatibility.

Streams LLM responses token-by-token back to Discord by editing a single message repeatedly as new tokens arrive, creating a live-updating effect rather than waiting for full completion. Implements a token buffer that batches tokens into chunks (typically 50-100 tokens) to avoid hitting Discord's message edit rate limit (5 edits per 5 seconds), with fallback to pagination if response exceeds 2000 characters. Uses Discord's message edit API with exponential backoff for rate limit handling.

Enforces permission rules by checking Discord user roles before executing commands, with optional per-user or per-command token budgets to prevent abuse or runaway costs. Implements a quota tracking system (in-memory or database-backed) that counts tokens consumed per user per day/week/month, blocking requests that exceed limits with a user-friendly error message. Integrates with Discord's role system to map roles to permission tiers (e.g., 'supporter' role gets 1000 tokens/day, 'admin' gets unlimited).