Compass

Inspiration

Most AI tools treat every conversation the same they respond, but they don't hold context. We watched teams use ChatGPT for project work and noticed the same pattern: great first answer, then total chaos by message 10. The plan drifts, the AI forgets the goal, and nobody catches it. We wanted to build something that actually manages the workflow not just the response.

What it does

Compass is a context-aware AI collaborator that keeps any project on track — for anyone, from engineers to finance teams to marketers. You define a goal, Compass builds a structured plan and immediately starts doing the work step by step. The key innovation is drift classification: every message is analyzed before the AI responds to determine if you're on-track, accidentally going off-topic, or intentionally changing direction. Accidental drift gets a gentle redirect. Intentional pivots trigger a **plan update mode — showing exactly which steps are affected and offering three options: modify the plan, insert a step, or branch into a parallel plan. Every message also gets a real-time safety score that flags PII, over-reliance risks, and transparency issues before they reach the AI.

How we built it

Lovable for the full-stack prototype. Gemini Model(gemini-flash-2.5) powers three separate AI calls per message: drift classification (returns structured JSON), response generation, and pivot impact analysis. The safety layer is pure regex, no AI needed, ported from a custom pattern library covering PII detection, over-reliance signals on financial/medical/legal domains, and transparency checks.

Challenges we ran into

The hardest problem was the line between unintentional drift and intentional pivot - they look identical at the message level. We solved it with a classification prompt that defaults to the more conservative label when uncertain, plus an escalation rule: two consecutive off-topic messages automatically promote to pivot treatment. The second challenge was Streamlit's rerun model - every interaction rerenders the full script, so we had to be careful about when API calls fire and how state persists across phases.

Accomplishments that we're proud of

The drift classification works reliably across completely different domains — the same system handles a software project, a marketing campaign, and a budget forecast without any domain-specific tuning. The plan branching UX is also something we're proud of: you never lose your original plan, you just fork it, which mirrors how real projects actually evolve.

What we learned

Prompt engineering for structured JSON output is an underrated skill. Getting the classification prompt to consistently return parseable JSON - especially handling edge cases like short acknowledgments and ambiguous pivots took more iteration than the UI. We also learned that the most valuable AI feature isn't intelligence, it's memory of intent.

What's next for Compass

Persistent sessions across browser reloads, team collaboration (multiple users on one plan), and deeper Tavily integration, so research-heavy steps automatically pull real-time data. Longer term: async agents that work on steps in the background while you review the last output, turning Compass from a turn-by-turn collaborator into a true parallel workstream manager.