AI Engineering — Case Study
Trackr — Agentic Health Tracking & Nutritional Intelligence
An AI agent that lets you log a workout or a meal the way you'd tell a friend about it — then remembers it, looks up the nutrition, and tracks your progress over time.
01 — Overview
Logging health data shouldn't feel like data entry
Most fitness and nutrition apps ask users to fill in structured forms — pick a food from a database, enter grams, select an exercise from a list. Trackr instead takes natural language: "had two eggs and toast, then ran 5k" — and does the structuring itself.
The harder problem isn't parsing that sentence once — it's turning a stream of casual, inconsistent daily entries into a reliable longitudinal record the user (and the agent) can actually reason about weeks later.
02 — How it works
Tool-calling + persistent memory
Natural-language logging
Users describe meals and workouts conversationally. The agent extracts structured entries (food items, quantities, exercise type, duration/intensity) from free text.
Tool-calling integrations
Real-time nutritional lookups are handled through tool calls rather than relying on the model's own memorized nutrition data — keeping figures grounded and current.
Persistent memory layer
Logged entries feed a longitudinal store the agent can query, enabling trend-aware answers ("you've been under your protein target most of this week") instead of one-off responses.
That memory layer is also what powers personalized recommendations — because the agent isn't just reacting to the latest message, it's reasoning over the user's actual history.
03 — Tech Stack
What it's built with
04 — Learnings
Grounding beats memorized facts
Routing nutritional lookups through a real tool call, instead of trusting the model's internal knowledge, avoided the plausible-but-wrong numbers that undermine trust in a tracking tool.
Memory design is a UX decision, not just infra
What the agent chooses to remember and resurface shapes how "personalized" it feels — this mattered more to perceived quality than raw model capability.
Natural language input needs a forgiving parser
Real users describe meals inconsistently. Handling ambiguity gracefully (asking a clarifying question rather than guessing silently) improved logging accuracy.