Natiq

Inspiration

As a computer scientist raised by a woman who's worked in real estate her entire life from call center to finishing degrees while pregnant and raising her kids during all of this and ending up as the director of land and real estate in the country. I thought to use an ai self improving voice agent to enhance one of the cornerstones of brokerage and estate which starts at the call center.

What It Does

Natiq is a self improving voice sales agent for real estate that handles live lead qualification calls, analyzes each conversation afterward, identifies what worked and what failed, detects objection and engagement patterns, and automatically updates its strategy to improve conversion rates. Instead of running on a static prompt like traditional voice agents, Natiq continuously learns from both successful and failed calls, evolving its conversational policy to increase bookings and reduce lost leads over time.

How We Built It

We built Natiq as a closed loop learning system with five stages: Call, Analyze, Learn, Improve, and Deploy. The agent handles a live call, records the transcript and outcome, then uses GPT-4o to extract structured insights such as objections, key phrases, and conversion signals. These insights are aggregated into recurring success and failure patterns with confidence scoring, stored in Supabase, and periodically used to update a versioned sales strategy. The improved strategy is then deployed automatically for future calls, creating a continuous optimization cycle using ElevenLabs for voice execution and a FastAPI backend for orchestration.

Challenges We Ran Into

One of the biggest challenges was avoiding overengineering the system within hackathon time constraints while still demonstrating real self improvement. Designing a learning loop that felt credible without building full reinforcement learning infrastructure required careful structuring of pattern detection, confidence scoring, and strategy updates. We also had to manage prompt drift, ensure version control for strategies, and keep the architecture simple enough to deploy quickly while still being scalable.

Accomplishments That We’re Proud Of

We successfully built a working end-to-end pipeline where each call produces structured learnings, updates pattern confidence, and can trigger automatic strategy upgrades. We implemented version tracking for conversational policies, dynamic prompt assembly based on high confidence patterns, and measurable conversion tracking per version. Most importantly, we demonstrated that a voice agent can meaningfully evolve based on real call outcomes rather than remaining static.

What We Learned

We learned that building a self improving system is less about throwing bigger models at the problem and more about building a loop that actually listens. The breakthrough was not some exotic AI trick, it was disciplined feedback. Clean data, clear conversion metrics, and versioned strategy updates mattered more than adding another model or fancy layer. At some point we realized we were not training intelligence, we were training accountability. Every call had to answer one question did this help us close or not. We also learned that picking a single vertical like real estate was not limiting, it was liberating. When the goal is specific and measurable, improvement becomes obvious. Trying to build a generic voice agent would have been impressive. Building one that actually closes better each week is far more satisfying.

What’s Next for Natiq

Next, we plan to expand Natiq beyond simulated calls into real brokerage pilots, integrate deeper CRM connectivity, and implement more granular performance analytics such as objection-level success rates and A/B testing across strategy versions. Long term, we aim to generalize the framework to other high-value sales industries while keeping the core principle intact: every conversation should make the system smarter.