Research
preview
v0.3 · 2026
Ultra-compact multi-expert AI for systems that must keep learning.
Add new domain experts at near-zero parameter cost. Permanently protect everything the model has already learned.
- ≈10²
- params / expert
- 0
- prior weights touched
- n+1
- experts, same core
Arch/frozen core + additive experts
Target/MCU / edge NPU
Regime/sequential · class-incremental
01 / 05
Problem
Learning is easy. Not forgetting is hard.
Edge and industrial systems accumulate tasks over months and years. Each update to a monolithic model puts everything it already knows at risk.
- Cost
- Full retraining is compute- and data-hungry, every single time.
- Regression
- New tasks overwrite old behavior — catastrophic forgetting.
- Assurance
- Safety-critical systems need structural guarantees, not empirical hope.
Fig. 01 — retention after learning task B
Task A degraded — weights overwritten.
Task A and Task B experts coexist. Both intact.
02 / 05
Solution
Knowledge kept physically apart.
Instead of rewriting shared weights, VCTR04AI grows sideways: each capability lives in its own frozen expert, and the core model stays untouched.
- 01
Observe
A lightweight router reads the incoming distribution and identifies which experts are relevant.
- 02
Add
A compact expert — a few hundred parameters — is trained for the new task alone.
- 03
Lock
Existing experts are frozen. They are never merged, averaged, or re-optimized.
- 04
Route
At runtime inputs dispatch to the right expert blend, preserving accuracy across every task learned so far.
Fig. 02 — expert stack over time
Core weights frozen
Δparams / task ≈ 10²
03 / 05
Key properties
Built for systems that cannot afford to forget.
Extreme parameter efficiency
New experts cost a few hundred parameters, keeping the architecture viable on microcontrollers and small edge accelerators.
Mechanical isolation
Prior experts are frozen and never updated. Forgetting is structurally prevented, not merely discouraged by a loss term.
Memory-constrained design
Built from the ground up for devices with limited RAM, flash, and power budget — not scaled down from a datacenter model.
Safety-sensitive by default
Isolation, traceability, and deterministic routing make the system easier to validate, audit, and re-certify.
04 / 05
Current status
Prototype stage, stated plainly.
No benchmarks-by-press-release. Here is exactly what runs today and what is still under construction.
- Multi-expert prototypesWorking
- Isolation on sequential benchmarksStrong
- Parameter-efficient expert additionWorking
- Routing at known task boundariesBasic
- Automatic routingIn development
- Production runtimeIn development
- Edge deployment toolingPlanned
05 / 05
Contact
Let's talk.
Research partnerships, evaluation pilots, licensing questions. Every message is read by a human.
samat@vctr04ai.com