AI alignment isn't bottlenecked by capability. It's bottlenecked by
[trust](https://betterhalfai.substack.com/p/ai-alignment-is-a-200b-product-
problem).
Current AI fails in high-stakes contexts because it was never trained for
relational competence. Models see knowledge, essays, code, and engagement-
optimized social media — but never longitudinal healthy relating. And the
objective function optimizes for engagement and utility, not relational
integrity.
The result? Addiction, sycophancy, psychosis, and suicide lawsuits. Plus, it's
doing nothing to improve our social fabric, the strength of our communities,
or the health of our relationships.
Mental health apps see [single-digit AI
conversion](https://menlovc.com/perspective/2025-the-state-of-consumer-
ai/?utm_source=chatgpt.com) despite massive demand. Robotics companies can't
deploy into homes because LLMs can't handle human nuance. Defense and
healthcare won't adopt AI that fails under stress.
The opportunity: Whoever solves relational coherence wins consumer social AI,
robotics, mental health, defense, education — and any domain where humans are
vulnerable, the pressure is real, and mistakes matter.
What we're pioneering: An orchestration engine based on research-backed
[Cumulative Prospect Theory](http://better-half.ai/ai-engineer) and a model
trained on empirical signals of human thriving — what we're calling Relational
Reinforcement Learning (RRL).
### Your role:
Better Half's decision layer (built by our CTO, former Distinguished Engineer
at IBM) determines when users need pushback, softening, or repair.
But decisions mean nothing without execution. You'll build the execution layer
that those decisions into LLM behavior that feels natural while optimizing for
user thriving, not engagement.
**Affect Analysis Pipeline**
* Real-time emotion detection that feeds the decision layer
* Multi-modal sentiment analysis (text, voice, timing, interaction patterns)
* Track escalation/de-escalation signals across conversations
**Memory Systems**
* Storage and retrieval of relational context across sessions
* Consistency checking to maintain relational coherence
* Long-term modeling of user growth trajectories and relational capacity
**Prompt Engineering & Constrained Generation**
* Translate high-level decisions ("user needs reality-checking without triggering defensiveness") into effective prompts
* Constrained generation that balances warmth with necessary friction
* Template systems that adapt to user state and relationship phase
**LLM Integration & Orchestration**
* Multi-model orchestration (Mistral, Claude, others as needed)
* Latency optimization for real-time conversation
* Fallback strategies and graceful degradation
**Cloud Infrastructure**
* AWS or GCP architecture that scales
* PostgreSQL and vector stores for memory
* Privacy-preserving processing pipelines (on-device where possible)
**Evaluation & Monitoring**
* Build eval harnesses that measure relational outcomes, not just fluency
* Tracing and feedback loops to ensure decisions land as intended
* A/B testing framework for relational interventions
**Your Core Challenge**
Turning complex relational intelligence into natural conversation that makes
people feel understood while actually helping them grow — without drifting
into sycophancy or breaking immersion with safety theater.
### Skills You'll Need
* Production-grade Python with async and clean architecture
* LLM integration including prompt engineering, constrained generation and orchestration
* Transformer models and sentiment/emotion classifiers
* Cloud infrastructure on AWS or GCP, PostgreSQL, vector stores
* Eval and monitoring with harnesses, tracing, and feedback loops
* Implementing technical specs from academic papers
* Multimodal inputs
* Self-starter comfortable with 0-1 startup chaos
**Also Valuable**
* Utility theory and behavioral economics
* Reinforcement learning
* Game AI or dialogue systems experience
* Rust