Technology at DREAM

Human-Aware AI, Ethically Engineered

At DREAM, we don’t just use artificial intelligence—we investigate it, reimagine it, and rebuild it for something bigger: human wellbeing.

This section is where we pull back the curtain on the research and technical thinking that powers DREAM. If you’re curious about how AI and machine learning actually work—and how we’re reshaping them for mental health support—you’re in the right place.

What Kinds of AI Do We Work With?

We explore multiple kinds of AI models, each with its own strengths:

• LLMs (Large Language Models): These are the big brains behind most of today's advanced AI systems. Trained on billions of words from books, articles, websites, and more, LLMs learn to predict and generate human-like language. We use LLMs as a base layer in DREAM, then fine-tune their behavior for emotional safety, clarity, and trust.

• SLIM (Sparse Logic Inference Models): Developed right here at DREAM, SLIM is our ethics-first reasoning engine. Unlike LLMs, which rely on statistical patterns in language, SLIM is rule-based and deterministic. It brings structure, consistency, and clarity to decisions—especially when values or safety are on the line.

• Multimodal Models: We’re also working on models that don’t just process text, but combine voice, tone, timing, and other biometric signals. Our VEIL system (Vectorized Emotional Inference Layer) is one example—it listens not just to what is said, but how, helping DREAM respond with greater emotional intelligence

• Reinforcement + Human Feedback Loops: We explore ways to train AI with human values directly in the loop—through ongoing feedback, testing, and ethical review.

How Does AI Learn, Anyway?

Training AI is a bit like raising a child—with massive amounts of data and lots of trial and error.

1. Data Ingestion: AI systems start by absorbing data—everything from books and websites to clinical guidelines and anonymous conversations (when ethically sourced). This creates a vast map of patterns in language, logic, and behavior.

2. Training Pipelines: That data flows through a training pipeline where models learn to predict, classify, or respond to inputs. They adjust billions of internal parameters to match expected outputs—like predicting the next word in a sentence or deciding if a response is calming or harmful.

3. Fine-Tuning & Safety Layers: Once trained, we fine-tune models for our specific needs—mental wellness, trauma support, empathy, and trust. This includes filtering harmful outputs, adjusting tone, and injecting ethical reasoning.

4. Memory & Personalization: We also experiment with “memory” systems that allow the AI to recall prior context, personal history, or ongoing emotional states—without compromising privacy or consent.

Sparse Logic Inference Modelling (SLIM)

Want to go deeper? Our research paper lays out the SLIM paradigm, contrasting it with conventional AI models and outlining how sparse logic, value prioritization, and trauma-informed design can work in harmony to build systems that truly support people — not just analyze them.

DREAM's Programming Directives

SLIM is our ethical decision-making engine. Instead of chasing performance metrics, SLIM prioritizes transparency, integrity, and human dignity — a radical shift away from black-box AI systems. It’s designed to think clearly, act compassionately, and hold nuance — even under stress.

Vectorized Emotional Inference Layer (VEIL )

In future development stages, DREAM's MEMORY SYSTEMS will evolve into a multi-modal, time-aware system that integrates biometric signal processing - through voice data— as a core input for emotional understanding.