Research
Deep dives, engineering notes, model cards, and points of view from the team building Skrrol and Unitly. Analyticity Tech is an applied AI research and product company; this page is the index for everything we publish on generative-media provenance, consent-verified voice cloning, unit-aware numerical reasoning, evaluation methodology, and responsible deployment.
C2PA credentials, verification workflows, platform handling.
Voice cloning consent, identity verification, revocation.
Unit-aware math, fail-loud engineering, audit-grade outputs.
Versioned eval harness, model cards, reproducible claims.
Six operating principles, policy posture, takedown flow.
Research · Apr 9, 2026
We describe the architecture, training setup, and evaluation suite behind Skrrol v2 — a multimodal model for short-form video reasoning, engineered around a sub-second latency envelope for production workflows.
Large language models routinely drop, confuse, or invent units when reasoning over numerical problems. We show a reproducible failure mode, introduce a benchmark, and describe the inference-time approach Unitly uses to close the gap.
A walkthrough of the evaluation pipeline we use across every Analyticity model — continuous, versioned, adversarial, and cheap enough to run on every pull request.
Capabilities, training data, intended use, known limitations, red-team findings, and deployment considerations for Skrrol v2.
The interesting work in 2026 is not the frontier model — it is the system around it. Data contracts, evaluation, latency, trust, and governance are where applied AI will be won. Our point of view.