PromethAI

Tracks user progress across nutrition and arbitrary personal goals by accepting periodic user input (food logs, workout data, habit completion) and using an LLM agent to analyze trends, identify patterns, and generate contextual insights. The system maintains goal state across sessions and uses the LLM to reason about progress relative to user-defined targets, enabling adaptive feedback without hardcoded rule engines.

Parses free-form user nutrition input (e.g., 'had 2 eggs, toast, and coffee') using LLM-powered natural language understanding to extract food items, quantities, and estimated macronutrients. The system normalizes extracted data into a canonical format (calories, protein, carbs, fats) and optionally cross-references a nutrition database to improve accuracy, enabling users to log meals conversationally without structured forms.

Accepts high-level user goals (e.g., 'lose 10 pounds in 3 months') and uses an LLM agent to decompose them into actionable sub-goals and daily tasks with specific metrics. The agent reasons about goal feasibility, identifies dependencies between tasks, and generates a prioritized plan that the user can execute incrementally. The system maintains the plan state and adjusts it based on progress feedback.

Maintains user state across multiple conversation sessions by storing goal definitions, progress history, and previous LLM interactions in a persistent backend. The system retrieves relevant context when the user returns and injects it into new LLM prompts, enabling the agent to provide continuous, contextual feedback without requiring users to re-explain their goals or history.

Analyzes user progress data over time (nutrition logs, goal completion rates, habit streaks) and uses an LLM agent to generate contextual, personalized feedback that adapts to detected patterns. The system identifies trends (e.g., weekend diet slips, morning consistency) and generates targeted recommendations without requiring explicit rule configuration, enabling dynamic coaching that evolves with user behavior.

Abstracts LLM provider selection (OpenAI, Anthropic, Ollama, local models) behind a unified interface, enabling runtime provider switching based on cost, latency, or availability constraints. The system implements fallback logic (e.g., use Anthropic if OpenAI quota is exhausted) and cost-aware routing (e.g., use cheaper models for simple tasks, expensive models for complex reasoning), reducing operational costs and improving resilience.

Engages users in multi-turn conversations to refine vague or ambiguous goals through LLM-driven clarification questions. The agent asks targeted questions about constraints, timelines, and success metrics, then iteratively updates the goal definition based on user responses. This reduces friction in goal setup and ensures the system understands user intent before generating plans.

Aggregates multi-dimensional progress data (nutrition metrics, habit completion, goal milestones) into unified dashboards and visualizations. The system computes derived metrics (weekly averages, trend lines, streak counts) and formats them for display, enabling users to see progress at multiple time scales without manual calculation.

Captures explicit user feedback on LLM-generated recommendations and insights (e.g., 'this feedback was helpful', 'this plan is unrealistic') and stores it for analysis. The system can use this feedback to identify failure modes, retrain or fine-tune models, and improve future recommendations without requiring manual annotation.