One Person, Many Voices: How to Build Agents That Sound Like a Real Person — Ben Carsley

Ben Carsley, a founding engineer at The Sentience Company, joins Angelina Yang on TwoSetAI for a 62-minute workshop on personalized writing models.

One person writes in several voices

A register is a context-specific style: messages to friends, work Slack, professional email, and academic writing all differ.
Recipient relationships also matter; a close colleague may get the same casual voice as a friend.
Ordinary text embeddings group passages by subject; style-focused embeddings help group them by how someone writes.
Sentence length, punctuation, and emoji frequency also provide inexpensive style measurements.

Sentience first clustered users’ writing and produced a guide plus representative examples for each register.
Unsupervised clustering, which finds groups without predefined labels, even surfaced Carsley’s sarcasm across different settings.
Accumulating instructions creates a growing burden on the agent’s context, the information available while it answers.

An internal Norman Cousins prototype used prompting and retrieved writing samples to imitate the author.
Asked for a React component about the mind–body problem, it happily produced code.
Carsley uses that failure to motivate post-training: changing learned parameters to specialize behavior.
Relationship memory still supplies facts and context; personalized weights and retrieval serve different jobs.

LoRA, low-rank adaptation, trains a compact set of parameter updates while keeping a shared base model frozen.
A serving system can select each user’s adapter while sharing GPU capacity; Carsley cites vLLM, an open-source inference server.
His six-user example illustrates shared serving, rather than reporting a measured capacity limit.
He provides no training-cost breakdown or production benchmark establishing the total economics.

Carsley says users can export or delete their personal adapter, stored separately in private cloud storage.
The personal-model feature is described as an internal cohort requiring connected iMessage and Gmail accounts for enough training data.
Captured audio cannot yet train the writing-style system; speaker identification and voice cloning remain separate work.
A colleague’s automated web-client build still needed intervention when an agent began testing in production.

Carsley presents side-by-side writing samples and says personalized outputs sometimes pass an AI-writing detector.
These demonstrations offer a design direction; detector results do not establish writing quality or reliable identity imitation.
Caveat: quality, privacy, and exportability are the vendor’s claims; the workshop supplies no independent audit or controlled comparison.

“One person contains many voices.”