Skip to main content
ROCSI
ROEN
Menu

Current Work / 04 / MVP in development

MVP in development
Domain
Applied AI / web infrastructure
Started
2026
Development
Prototype and offline evaluation
Environment
Local, no public deployment

ROCSI AI Assistant

Answers grounded in an organization's approved information.

ROCSI AI Assistant is a B2B product in development: a website-embeddable assistant designed to help visitors discover relevant information, inspect its sources, and understand the limits of an answer. The current prototype runs locally; it is not a public service and does not yet use a generative language model.

All Current Work →

01 / Why

A website publishes information. Visitors need to find it.

The problem

On many organizational websites, common answers are scattered across pages, documents, and hard-to-navigate sections.

An assistant that invents information or mixes sources across customers would create new problems rather than reduce friction.

Why ROCSI is working on it

We are testing whether a portable assistant can surface each organization's approved information and point to the source used, without claiming knowledge it cannot substantiate.

ROCSI and the Connector for NMKR website are the initial evaluation contexts, not commercial installations.

02 / System

From approved source to justified answer.

The diagram shows the intended conceptual flow. Prototype stages are not equivalent to a launched generative RAG service.

  1. 01

    Website

    A site embeds an isolated widget, with separate per-customer configuration.

  2. 02

    Approved sources

    Candidate pages and documents require access controls, review, and approval before activation.

  3. 03

    Search and evidence

    The prototype evaluates RO/EN lexical matching and citation of authentic passages.

  4. 04

    Interaction

    The product direction includes bounded answers, justified refusals, and a human contact path when needed.

03 / Technical state

What works locally, and what remains to be proven.

Verified state as of 5 November 2026, MVP-05. No corpus activation or public demo has been approved.

Local prototype

  • JavaScript widget with isolated interface, separate configurations, and local ROCSI and NMKR demos.
  • Prototype API with deterministic responses and refusal when information cannot be supported.
  • Regression tests and checks for source references and cross-customer separation.

In validation

  • Quality of Romanian and English retrieval, including new questions and paraphrases.
  • Review and exact approval of corpora before activation.
  • Controlled source expansion and repeatable onboarding of organizations.

Future direction

  • Language model integration and RAG answers, subject to separate evaluation.
  • Multi-customer administration, feedback, and consent-based contact features.
  • Pilots and commercial offering only after quality and security criteria are met.

04 / Principles

Boundaries and provenance are part of the product.

01

Controlled sources

Each organization determines what information may support an answer.

02

Customer isolation

One customer's information must not appear in another's answer.

03

Justified refusal

Missing evidence must not be replaced with speculation.

04

Bilingual evaluation

Romanian and English quality is measured separately, including on unseen questions.

05

Human gates

Corpus publication, model use, and launch require separate decisions.

05 / Research relationships

Applied AI, trustworthy information, and reusable infrastructure.

Research becomes testable engineering here, rather than broad claims about AI.

AI & Applied Intelligence

Evaluates when bounded AI assistance can use controlled information with measurable results.

Trust & Information Integrity

Studies provenance, citations, and refusal when evidence is missing.

Open Digital Infrastructure

Designs a portable widget and reusable architecture across different websites.

Explore ROCSI research →

06 / Project state

MVP in development

Active development / not launched

The offline MVP-05 evaluation has not yet passed the quality thresholds required to activate corpus-based retrieval. The safeguards tested so far are not a universal guarantee. There is no public service, customer onboarding, or live LLM integration.

Current focus

  • Improving retrieval, especially for natural Romanian formulations.
  • Independently validating new RO/EN sources and outcomes.
  • Keeping separate gates for corpus activation, LLM integration, and deployment.

07 / Next

Technical milestones, not promised commercial dates.

This page is updated after validated milestones, not on every code change.

  1. 01Controlled expansion of candidate documentation and bilingual evaluation.
  2. 02Testing a repeatable website onboarding process.
  3. 03Evaluating generative RAG, privacy, and costs before a pilot.

08 / Follow / collaborate

A project made visible, but not yet offered as a service.

This page is the stable public reference for the project's progress. There is currently no public demo or installation offering.

All Current Work →Workbench →Discuss collaboration →