BabydollAI

🎭 BabydollAI: Preventative Behavioral Onboarding Middleware

1. The Inspiration

Modern dating applications are facing a silent trust and safety crisis: the weaponization of wellness culture. High-risk, avoidant, and manipulative individuals have adapted to modern psychology trends. Instead of using obvious red-flag language, they use contemporary terms like "protecting my peace," "energetic alignment," or "unresolved drama" to justify emotional neglect, evade mutual accountability, and gaslight partners. Current dating platforms offer no preventative shields against this sophisticated, trends-based behavior. BabydollAI was built to protect users from this hidden emotional toxicity at the very front gate of platform onboarding.

2. How We Built It

The application acts as a preventative, behavioral middleware gateway engineered directly into dating application onboarding flows. Instead of asking standard, superficial profile prompts, it evaluates users using subtle behavioral situational triggers.

The underlying technology stack includes:

• Development Tooling: The system architecture, syntax logic, and multi-inference processing pipelines were entirely engineered and generated utilizing the OpenAI Codex framework.
• Official Enterprise Inference: The primary model framework natively targets the OpenAI GPT-5.6 production engine, utilizing strict JSON schema enforcement to pass structured psychological metadata back to host platforms.
• Public Fallback Engine: Deployed utilizing an integrated fallback pipeline to the ultra-low-latency Groq Llama-3.1-8b-instant infrastructure, allowing rapid public evaluation without API cost friction.
• Frontend UI: A responsive, production-ready dashboard built using Streamlit (Python).

3. Challenges We Faced & The Engineering Breakthrough

During development, we uncovered a critical technical bottleneck: standard AI models are easily fooled by therapy-speak. Because base large language models are heavily trained on polite textbook data, initial testing scored manipulative text as "90% secure" simply because responses contained words like "inner work."

We successfully solved this by implementing strict forensic constraints within our system instructions, forcing the engine to analyze the contextual mechanics of linguistic accountability rather than individual keywords. This calibration successfully transformed BabydollAI into an objective, un-foolable gatekeeper.

To track this mathematically, we modeled a weighted behavioral composite index, calculating the final relational safety score $S_{total}$ using the following formula:

$$S_{total} = w_1 \cdot S_{secure} - (w_2 \cdot A_{avoidant} + w_3 \cdot T_{toxic}) \cdot (1 + \delta \cdot D_{deception})$$

Where:

• $S_{secure}$, $A_{avoidant}$, and $T_{toxic}$ represent the underlying base sub-ratings scored by the forensic parser.
• $w_1, w_2, w_3$ are normalized weight coefficients satisfying $\sum w_i = 1$.
• $D_{deception}$ represents the detected percentage of weaponized therapy buzzwords, scaled by a dampening penalty factor $\delta$.

4. What We Learned

Building BabydollAI proved that linguistic patterns are incredibly accurate predictive indicators of adult attachment styles and behavioral accountability. We learned that to build effective modern safety infrastructure, AI systems cannot rely on static keyword matching or surface-level sentiment analysis. True preventative security requires prompt-engineered contextual modeling capable of interpreting the structural psychology behind conversational intent.

Built With

• github
• gpt-5.6
• groq
• json
• llma3.1
• openaiapi
• openiacodex
• python
• streamlit