Halo: The Integrated AI Mediator & Relationship Coach

The Judge of facts. The Coach of hearts. The Mediator who never misses a beat.

Comment

The Inspiration: The "He Said/She Said" Trap

In many relationships, the greatest wound isn’t the argument itself—it is the erosion of reality. Every couple falls into the "Subjective Memory Trap," but for those experiencing gaslighting or emotional manipulation, this trap becomes a weapon. When "flight or fight" kicks in, objective facts are often replaced by a rewritten history that leaves one partner feeling silenced and insane. We were inspired to build HALO to solve this fundamental human vulnerability.

We envisioned a "Third Entity" in the room—not a spy, but a neutral, incorruptible guardian. By leveraging the hardware we already carry—iPhones, Apple Watches, and Google Nest Cams. It provides a source of truth for those whose reality is being questioned, ensuring that "what happened" is never a matter of who can shout the loudest.

Also, by establishing a digital 'Household Constitution,' HALO transforms conflict into clarity. After a dispute, the system analyzes the interaction against the couple’s agreed-upon values, providing an objective debrief that ensures both partners remain aligned with their shared values and move forward with restored accountability.

Beyond settling active disputes, HALO functions as a proactive Relationship Coach that identifies the subtle, non-verbal precursors to conflict. By recognizing one partner’s escalating behavioral patterns or environmental triggers in real-time, the system issues an intuitive prompt to the other partner. This nudge encourages immediate supportive action or collaborative de-escalation, allowing the tension to be diffused through partnership before it reaches a breaking point.

How We Built It: The Multimodal Brain

1. Technical Diagram Outline:

1) Entities:
Users (Partner A, Partner B)
Hardware (iPhone, Apple Watch, Google Nest Cam)

2) Data Flow:
Hardware → Mobile App (HealthKit/Camera Stream)
Mobile App → On-Device Filter (Face Blur/Privacy Shield)
Anonymized Data → Gemini 3 API (Reasoning/Coaching)
Gemini → Vector DB (Long-term Memory)
Alerts/Nudges → Apple Watch/Nest Speaker (Action)

3) Deployment & Interaction
iPhone: Maintains an acoustic "Truth Buffer" and acts as the manual summon button.
Nest Cam: Provides visual context to resolve "He Said/She Said" deadlocks.
Apple Watch (optional): Monitors health data to provide silent, proactive de-escalation nudges.

Summon HALO on your iPhone or type in the chat box with prompts like:

"Halo, who was responsible for this according to our Constitution?"
"Halo, playback the facts: what was actually said ten minutes ago?"
"Halo, help me articulate the pattern that led to this frustration."

2. HALO was built using Gemini 3 in Google AI Studio, leveraging its massive context window to act as a long-term "Relationship Ledger."

1) The Sensing Layer: We integrated HealthKit data for biometric stress detection. By monitoring Heart Rate Variability ($HRV$), HALO detects "Physiological Flooding" before a word is even spoken.

2) The Privacy Shield: We implemented a MediaPipe local filter. Faces of guests are blurred in real-time on-device, ensuring the AI only "sees" the behavior and context, never the identity of outsiders.

3) The Memory Decay Pipeline: To solve the privacy-vs-utility dilemma, we built a tiered storage system:
Tier 1: Raw metadata for 48 hours.
Tier 2: Semantic summaries for 30 days.
Tier 3: Irreversible Vector Embeddings for long-term pattern coaching.

The Challenges: Balancing Proactivity with Privacy

The biggest challenge was the "Creepiness Factor." Nobody wants a judge in their living room. We had to iterate deeply on the Circuit Breaker logic.

We learned that for an AI to be a successful coach, it must be Proactive, not just Reactive. If HALO only speaks up after a fight starts, it’s a nagging judge. If it sends a nudge to the husband’s Apple Watch when it senses his stress rising before he walks into the kitchen, it’s a savior.

What We Learned

We discovered that AI's greatest gift to relationships isn't its ability to talk; it's its ability to remember accurately and forget respectfully. By using Vector Embeddings, we proved that an AI can "learn" that a couple struggles every Sunday night without needing to store a single private recording of their Sunday dinner.

Built With

Share this project:

Updates