Under the hood
How EloDtx Works
Three systems. One purpose. Building real human connection at scale — with whispers, proximity, and a model that never stops learning.
The Whisper System
EloDtx never writes messages for users. It whispers suggestions.
The Philosophy
Human connection is irreplaceable. EloDtx's role is not to speak for you — it's to give you the clarity to speak better yourself. Every suggestion is framed as an observation, never a script. You decide what to do with it.
Example whisper
"She mentioned not going back to her hometown much — her answers suggest this means something to her. This could be worth exploring gently."
Ephemeral Processing
Message content is never stored. EloDtx reads a conversation context window in real time, extracts behavioral patterns, and immediately discards the raw text. Only the patterns remain — anonymised, aggregated, and tied to behavioral dimensions rather than anything you actually said.
I haven't been back to my hometown in years honestly…
That's interesting — what keeps you away?
She mentioned not going back to her hometown much — her answers suggest this means something to her. This could be worth exploring gently.
The Proximity Flow
From opt-in to conversation — every step privacy-first and consent-gated.
User enables proximity
With a single tap the user opts in. Proximity is always off by default — active consent is required every session.
Geohash location sent (±500m, never exact)
The device encodes the user's position into a geohash — a ~500m × 500m area code. Exact coordinates never leave the device.
EloDtx runs compatibility check (<2 seconds)
The engine cross-references all nearby opted-in users against the requesting user's 138-dimension behavioral profile in real time.
Apple Watch haptic tap
If a high-compatibility match is within range, a discreet haptic pulse fires on the wrist. No sound. No visible notification.
5-element alert card appears
The watch face shows compatibility score, shared traits, current activity, distance, and two action buttons.
Profile request sent
Tapping 'Connect' sends an anonymous interest signal. The other person receives an identical alert and chooses independently.
Consent → Messenger unlocks
Only when both parties tap 'Connect' does the messenger open. No one-sided reveals. No pressure.
The Behavior Learning Loop
Every interaction — active or passive — refines the model. The profile is never static.
Passive Signals
Profile138 dimensions
Active Signals
All behavioral signals are processed ephemerally. Message content is never stored — only anonymised patterns are retained to improve the model.
The Training Path
We're building our own model — one that has never existed before, trained exclusively on connection data.
Claude API Layer
EloDtx runs on Anthropic's Claude API with proprietary prompting, fine-tuned for connection intelligence. This gives us production-grade AI while we build our own.
Qwen 3.5 Fine-Tuning
Fine-tuning Qwen 3.5 on our proprietary behavioral dataset — 508 real user profiles, conversation patterns, and outcome data from the Baeyond POC.
Hybrid Deployment
Running our fine-tuned model alongside Claude for validation. Gradual traffic migration as our model proves equal or superior performance.
Full Proprietary EloDtx Model
Complete independence. A model trained exclusively on connection data, optimized with TurboQuant compression for 4.9x efficiency.
Claude API Layer
EloDtx runs on Anthropic's Claude API with proprietary prompting, fine-tuned for connection intelligence. This gives us production-grade AI while we build our own.
Qwen 3.5 Fine-Tuning
Fine-tuning Qwen 3.5 on our proprietary behavioral dataset — 508 real user profiles, conversation patterns, and outcome data from the Baeyond POC.
Hybrid Deployment
Running our fine-tuned model alongside Claude for validation. Gradual traffic migration as our model proves equal or superior performance.
Full Proprietary EloDtx Model
Complete independence. A model trained exclusively on connection data, optimized with TurboQuant compression for 4.9x efficiency.