Hey Internet

Accepts free-form text queries via SMS and routes them through an LLM inference pipeline that interprets intent from unstructured, often abbreviated mobile messaging syntax. The system handles SMS character limits (160-1600 chars depending on encoding) by chunking long queries and reconstructing context server-side, then returns responses formatted to fit SMS constraints with intelligent truncation or multi-message splitting.

Maintains conversation state across multiple SMS exchanges by storing message history server-side and reconstructing context from previous queries in the same thread. Uses phone number + timestamp-based message grouping to associate related queries, then injects prior exchange summaries into the LLM prompt to simulate multi-turn awareness without requiring explicit session management from the user.

Interprets natural language commands in SMS (e.g., 'remind me to call mom at 3pm', 'set a timer for 20 minutes', 'add milk to my shopping list') and translates them into executable actions via integration with device calendars, reminders, timers, and note-taking services. Uses intent classification to route commands to appropriate backend services (calendar API, reminder service, etc.) and returns confirmation via SMS.

Processes SMS queries that require real-time information (e.g., 'what's the weather', 'stock price of AAPL', 'nearest coffee shop') by routing them to web search APIs or structured data services, then synthesizing results into SMS-friendly summaries. Uses query classification to determine whether a response requires live data or can be answered from LLM training data, and applies result ranking/filtering to fit SMS character constraints.

Implements a freemium model where free-tier users receive a limited number of queries per day/month (likely 10-50 per day) before hitting rate limits, while paid users get unlimited or higher quotas. Uses phone number-based user identification to track usage, applies token-bucket or sliding-window rate limiting, and returns SMS notifications when limits are approached or exceeded.

Analyzes incoming SMS queries to classify intent (e.g., 'factual question', 'task creation', 'web search', 'calculation', 'creative writing') and routes them to appropriate backend handlers. Uses a lightweight classification model (likely fine-tuned LLM or rule-based heuristics) that runs server-side to determine which service should handle the query, enabling specialized handling for different query types without exposing complexity to the user.

Automatically formats LLM responses to fit SMS character constraints (160 characters for single SMS, or splits into multiple messages) while preserving readability and information density. Uses techniques like abbreviation expansion, emoji substitution, and intelligent truncation to maximize content within limits, and implements multi-message chaining with implicit continuation markers (e.g., '(1/3)') to signal multi-part responses.

Abstracts away carrier-specific SMS delivery by using a carrier-agnostic SMS gateway (likely Twilio, AWS SNS, or similar) to send and receive messages across all major carriers (Verizon, AT&T, T-Mobile, etc.). Handles carrier-specific quirks (e.g., message splitting, encoding differences, delivery delays) transparently, and provides basic delivery status tracking (sent, delivered, failed) via server-side logging.

Uses phone number as the primary user identifier (no email, username, or password required) to enable zero-friction signup and authentication. Stores minimal user data (phone number, usage quota, preferences) server-side and relies on SMS itself as the authentication mechanism — incoming SMS from a registered phone number is implicitly authenticated. Implements basic privacy controls (e.g., ability to delete account and associated message history).

Optimizes LLM inference latency for SMS use cases by using smaller, faster models (likely distilled or quantized versions of larger models) that prioritize response speed over maximum quality. Implements response streaming and early truncation — if a response is taking too long to generate, the system returns a partial response rather than timing out. Uses server-side caching of common queries to avoid redundant inference.