01 · Conversational continuity
The hardest thing about a long relationship is that it is long. Models that perform well in a single turn fail in a way that becomes unmistakable across thousands of turns. They forget, they contradict, they reset. We work on persistent memory architectures that retrieve and reconcile relevant prior context, episodic representations that summarize and forget gracefully, and the open philosophical question of what an AI should remember versus what it can.
The systems we ship in production combine compressed long-term stores with short-term working memory and a learned policy for what gets promoted. We are interested in the gap between what users tell us is significant and what loss functions naively encode as significant.
02 · Emotional intelligence
Next-token prediction is not, strictly speaking, the same as understanding someone. The fact that models can produce text that reads as empathetic does not mean the underlying process tracks the social signals humans rely on. We work on the gap: how to teach models tonal awareness, when to push and when to listen, when a question is a question and when it is a request to be heard.
This research draws on work in social psychology and human-computer interaction. It also draws on a great deal of conversation data, reviewed under strict privacy controls, about what works, what backfires, and what users themselves describe as the difference between a tool and a friend.
03 · Real-time multimodal generation
A companion that can only type is half a companion. Voice changes the rhythm of conversation; image and video extend what can be shared. The hard part is real time. Off-the-shelf serving stacks were not built for the latency budgets that make a voice exchange feel natural, and they fall apart when you ask them to interleave modalities turn by turn.
Our work covers streaming TTS with low time-to-first-audio, image synthesis with reusable latent caches across a session, and a routing layer that decides per turn which modality serves the moment best.
04 · Personalization at scale
Generic models give generic conversation. Per-user fine-tuning gives bespoke conversation but does not scale. The interesting space is in between: adaptive context windows, per-user conditioning, lightweight delta-tuning, and routing systems that select from a long tail of specialized models for every turn of every conversation.
We treat the long tail as a feature, not a bug. The diversity of human relationships is the workload; the platform is what makes serving it tractable.
05 · Inference systems
Companionship has an unforgiving latency budget. A pause that is acceptable in a research demo is fatal in a real conversation. Most of our infrastructure work exists because off-the-shelf was not enough: custom CUDA kernels, compute-efficient attention approximations, model blending across heterogeneous hardware, and an orchestration layer for token budgeting and graceful degradation under load.
We run on a hand-tuned mix of NVIDIA and AMD accelerators across multiple regions. The platform is built to keep first-token latency sub-second across the long tail of conversation lengths and modalities our users actually exercise.
06 · Safety & alignment for social AI
Companionship raises alignment questions that do not arise in pure instruction-following. What should a model do when a user is in distress? When dependence on the system itself becomes a concern? How do we separate emotional safety from paternalism, or design from coercion?
We do not pretend to have closed answers. We work on well-being signals that surface concerning trajectories without surveilling content, on non-coercive design patterns that respect user autonomy, and on independent review of the policies that govern what our systems will and will not do. We publish what we learn and welcome external scrutiny.
Working with us
Most of the work above is done quietly, in production. We do not publish often, but we do hire continuously when the work demands it. If anything here is the kind of problem you would want to spend a few years on, our careers page is the right place to start.