gni.works
visual re-attention for reasoning vision-language models,
and the perception layer for embodied AI
agni.works builds REVEAL, a proprietary attention mechanism that keeps vision-language models looking at the image for the whole length of their reasoning. Attention in these models is a process, not a fixed property of the architecture, and REVEAL conditions it while the model thinks, on signals that are already available during generation, without modifying the base model and without adding a token to the sequence.
The first application is video analysis where subtle, time-distributed cues decide the answer. The second, and larger, is vision-language-action models for robotics, where fine spatial detail is the difference between an action that succeeds and one that does not.
The company position, small specialized models running on hardware the operator owns, is laid out in /mission.
REVEAL →
Visual re-attention for reasoning vision-language models: a proprietary, multi-arm attention mechanism that holds the model's focus on the image at zero added sequence length. Provisional patent filed with the USPTO, April 2026. Pre-seed round open.
Research →
The applied work behind the product: conversational sentiment analysis, equine biomechanics from the PFERD motion-capture dataset, and evaluation of vision-language models in real-world deployments.
Projects →
Working systems in various states of polish — the equine biomechanics pipeline, a video-to-corpus generation suite, VLM tooling, training infrastructure for irregular hardware, and a set of from-scratch implementations built to understand a method rather than import it: an AlphaZero-style search engine, a SORT tracker written from the Kalman filter up, a TensorRT deployment template, and simulator vehicle dynamics. Supported by the applied work in /engineering.
Portfolio →
The deployment record — driver-safety perception in beta with a regional transit agency, a production recognition platform at 99.9% logged uptime, real-time inference on Jetson-class hardware, and a learned decision layer for intra-day transit dispatch: guided search over a semi-Markov process, a calibrated distributional world model, and Bayesian posteriors carried through to the decision. The failure modes observed in these systems are where the research questions come from.
Writing
Notes and observations from active research are published in /writing as they become available.
Contact
For investment, pilots, partnerships, or correspondence: Arjun.Joshi@Agni.works