Cross-model verification
Before a long, sourceless answer or an unfinished plan is treated as final, a genuinely different model reviews it — a real check, not the same model re-reading its own work.
AI Research
AI where it earns its place — not by default.
We study two things at once: where a verified, tool-using AI genuinely changes an outcome, and where a plainer, non-AI approach is the better engineering choice. Both are real research questions here, not marketing copy.
- A second AI checks the first, on purpose
- Cost-aware by design, not by accident
- Not every problem needs a neural network
Verified, not just generated
An answer that's checked before it's trusted.
This research underlies the AI Teachers assistant already running in classrooms — see it in AI in Education. Here, we focus on the open engineering questions behind it.
A stall watchdog, not a flat timer
Tracking real progress through a multi-step task means a quietly stuck attempt gets caught — a plain timeout would either fire too early on a hard task or too late on a broken one.
A real bug this caught
An early version of the verification step occasionally discarded a correct draft answer instead of approving it — traced to how the checking model was asked to confirm its own judgment, and fixed by changing that prompt, not by removing the check.
Two-tier, cost-aware by design
A fast, economical model handles most questions; a harder one is retried on a stronger model within the same request. The tier boundary itself is an open research question — set it too low and cost climbs, too high and easy answers get slower for no reason.
The other half of the research
Knowing when not to reach for a neural network.
Our own Facial Mimicry Cloning work is the clearest example: a real-time synthetic-video method built with no generative model at decode time at all — no GAN, no diffusion, no per-subject neural training. The result runs on an ordinary CPU with no GPU dependency, because the underlying problem turned out to be better solved by indexing and interpolating real recorded frames than by generating new ones.
That's a deliberate research stance: a system that reaches for AI only where it earns a real advantage in accuracy, cost or capability — and is honest about where a classical method wins instead.
Every allowed tool, in every AI system we build, is an explicit, short list — never "everything the model can imagine."
See this research applied in a real classroom.
The AI Teachers assistant is the applied side of this work.