veralocal.com
LocalFirst, truth seeking, agent driven, deterministic AI. Vera is designed to keep control with you: deterministic pipelines, clear inputs/outputs, and explanations you can audit.
A Local AI Agent is an AI that takes actions and completes tasks while running directly on your machine.
That is the core of Vera: research agents ingest your data, run deterministic logic, and produce auditable results in the browser—without sending your dataset to a cloud model by default. The domain veralocal.com reflects that commitment.
Vera runs on your computer. The core experience works in the browser with no server dependency.
Answers are built from ingested inputs so results stay evidence-based.
Composable agents that execute structured steps you can understand and reproduce.
Deterministic pipelines produce repeatable outcomes instead of stochastic guesses.
Import CSVs/TSVs and keep them in memory on your device.
Agents apply deterministic logic and domain skills to your dataset.
Results include the fields used and the intermediate values that drive the outcome.
Get grounded explanations of what changed and why.
veralocal.com is about where Vera runs: on your computer, with your data in memory for the core pipeline. Context is about what Vera runs on: a structured, auditable bundle—not a cloud prompt—made from the rows you ingested, explicit skill rules and assumptions, the agent’s computed results, and field wiring so Ask targets the right metrics.
Together, local + context means answers are computed from what you supplied on-device, instead of being stitched together by remote, open-ended text generation.
Locally ingested files become normalized in-memory structures you can query.
Explicit assumptions, formulas, and field catalogs—fixed rules, not “prompt vibes”.
Deterministic procedures that compute results and expose what was used.
A mapping layer so Ask can resolve “what you meant” to exact rows and fields.
“Local” alone does not guarantee trustworthy answers—you can still run opaque logic on-device. Vera pairs local execution with explicit context so results stay traceable: same inputs and configuration yield reproducible outputs, without token-metered cloud prompts for the core path.
Vera is built around a strict deterministic pipeline: ingest your data, normalize it into a structured in-memory store, compute agent-specific outputs using fixed rules and explicit assumptions, and then render explanations from those computed results.
Skills are the contract: explicit assumptions, formulas, and field catalogs that agents use at runtime. This keeps policy out of ad-hoc prompt text and makes reasoning reproducible and auditable.
Given identical inputs (the same ingested dataset, the same prompt, and the same agent configuration), Vera produces fully reproducible outputs because the core response path contains no stochastic sampling or random generation steps.
Most “AI” runs in the cloud and bills by tokens. Vera runs on your computer for the core pipeline: no token metering, no remote prompts by default, and a smaller security surface for sensitive datasets.
Vera’s Context Model is a structured “evidence bundle” assembled from three things: (1) your locally ingested data, (2) explicit, fixed rules and assumptions registered by skills and agents, and (3) a query-time mapping layer (`queryContext`) that tells the system which fields/rows you actually mean.
This makes context auditable: the system can point to the exact fields and computations behind an answer, instead of guessing what a user “probably meant.”
Vera’s Ask interface resolves your question against the active Context Model: it parses intent, routes the query to the relevant rows/fields, and computes the answer from the same deterministic result data the agent produced.
If the required data is missing, Vera asks for data or clarification instead of inventing facts.
Large Language Models are typically trained to predict the next token in text. At inference time, they combine learned representations (embeddings + attention) to produce continuations. Without strong grounding, this often leads to confident but incorrect statements—i.e., hallucinations—because the model is optimizing for plausible text, not verified facts.
Even with retrieval, many LLM workflows still rely on probabilistic decoding (beam search, sampling, temperature), which can change phrasing and correctness across runs.
Vera computes. LLMs decode. Computation follows explicit steps and rules; decoding turns a learned distribution into text.
Vera routes questions into structured field/row access via `queryContext`, and applies skills (explicit rules) to compute answers grounded in your ingested data.
Classic ML (including LLMs) optimizes training-time objectives to generalize across examples. That’s powerful for pattern recognition, but it is not the same as “proving” answers from your dataset.
With local execution, your raw dataset doesn’t have to leave your machine to get core results. External APIs are opt-in and treated as inputs; they change numbers only when the upstream data changes.