Updated September 19, 2026. This essay originally argued that the model was the easy part. We have revised that framing to reflect the importance of model research alongside system engineering.
A useful AI system depends on what it can perceive, what it remembers, and what it does with that information. Weakness in any one of these can determine the outcome. A model that misses an object cannot reliably reason about it. An agent that acts on stale state can make a bad decision even when its perception is correct.
Computer vision and deep learning are central to this problem. Representations determine which distinctions a model can make; training objectives and data shape what it learns; inference choices determine whether it can operate within a practical latency and cost budget.
Context and memory still matter
A model working inside a company needs current records and permissions. Retrieval, entity resolution, and persistent operational state help provide that context. Event sourcing can preserve a record of changes and allow state to be reconstructed.
Those properties are useful, but an append-only log does not establish that its input is true. A stored observation can be wrong. An audit trail can faithfully record an incorrect decision. Systems need validation and evaluation in addition to persistence.
OmniManas applies event-sourced operational state to business workflows and permissioned agents. Saathi is a personal AI assistant in private preview, combining voice, visual reading, and personal context with an accessible, Hindi-first experience.
Be precise about world models
A record of business events and a predictive visual world model solve different problems. We use operational state for the former. In this architecture, world models refer to learned representations that can support prediction of how a scene or environment changes.
Moving from remembered observations to predictive state requires evidence. Can the representation track an object through occlusion? Does it recognize uncertainty when information becomes stale? Can a model predict a future state better than a simple temporal baseline? These are experimental questions.
Evaluation connects the layers
Multivon provides tools for evaluating answers, agent behavior, and document workflows. Its public studies describe a particular dataset, model, and protocol; they do not establish universal reliability.
The financial evidence study illustrates this distinction. Asking a model to produce answers and evidence locations together worsened answer quality under the tested settings. A design intended to improve accountability introduced another source of error.
That observation supports a narrower next experiment: compare joint generation with separately bound evidence while keeping the model and task set fixed. It does not justify a general claim that evidence makes models worse.
From representation to action
Our direction connects representation learning and controllable generation to agents whose outcomes can be checked. The technical report explains the architecture, an executable planning demonstration, established evaluation methods, and our published evidence study.
Model research and system engineering should inform each other. An experiment should reveal whether a failure calls for better visual representations, different training data, better state management, or a more reliable verification step. The goal is to find the intervention that improves the measured task.