Humanika: the new AI initiative by 150 Por Ciento
150 Por Ciento logo 150 Por Ciento logo

Project — Case study

Humanika: governed AI digital collaborators for education and corporate training

How 150 Por Ciento designed and built Humanika, a platform to design, integrate and operate AI digital collaborators connected to an organization's knowledge, systems, permissions and people.
Client
Humanika (150 Por Ciento)
Industry
AI · Education & corporate training
Tech stack
Claude (Anthropic), OpenAI, Google Gemini, self-hosted open models, RAG, PostgreSQL, pgvector, Model Context Protocol (MCP), FastAPI, Next.js, React, TypeScript, WhatsApp Business, Microsoft Teams, Moodle, Academusoft, HubSpot, Google Drive, Microsoft 365, SharePoint, OneDrive

The problem: AI that talks but cannot operate

Most organizations we work with already tried AI. A chatbot answered a list of scripted questions; a general-purpose assistant wrote good text but could not see the student record, the course catalog or the CRM. Neither could take an action under the institution’s own rules, and neither left a trail an auditor could follow.

In education and corporate training the gap is sharper: answers must come from approved regulations, course content and academic systems, and many interactions happen on WhatsApp rather than on a website. Humanika was built to close that gap without starting from zero for every new client.

Our approach: one digital collaborator, one job

Humanika does not sell a chatbot. It designs, integrates and operates digital collaborators: AI agents that perform a specific function inside an organization, such as an academic tutor, an admissions agent, a help desk, a corporate trainer or an internal knowledge assistant.

Each collaborator belongs to one client and has its own knowledge, its own authorized actions, its own channels and its own escalation rules. The work follows four steps: understand the challenge, organize the knowledge, design the collaborator, and measure and evolve. Collaborators are grouped into five solution lines: corporate training, organizational knowledge, service and support, sales and growth, and productivity and operations.

Architecture: the model reasons, the platform operates

The core principle is a clear split of responsibilities. The language model interprets intent, holds the conversation, consults knowledge and decides when it needs a tool. Humanika Core, the control plane, manages clients, collaborators, credentials, integrations, channels, conversations, usage and errors.

  • Knowledge: retrieval-augmented generation (RAG) on PostgreSQL with pgvector. Sources such as PDFs, URLs and Google Drive folders are synced per collaborator, never into a shared global index.
  • Integrations and capabilities: an integration is a connection (Moodle, Academusoft, HubSpot, Google Drive, Microsoft 365, SharePoint, OneDrive, a custom API); a capability is a single action on it, such as checking course progress or creating a lead. Moodle and Academusoft are also exposed through MCP servers.
  • Models: Claude-first, but behind a provider-agnostic layer that can route to Anthropic, OpenAI, Google Gemini or self-hosted open models.

Governance and security by design

Governance is not a layer added at the end; it decides what the collaborator is allowed to do.

  • Per-client data isolation: knowledge, conversations and credentials of one organization are never available to another.
  • Role-based access: the operations team works with Admin, Consultant, Operator and Viewer roles, and the backend enforces every permission.
  • Governed capabilities: the model can only request actions that were explicitly authorized for that collaborator; the platform validates, executes and audits each one.
  • Mandatory audit: configuration changes, access events and actions are logged.
  • Human supervision: conversations escalate to a person when the user asks for one, confidence is low, an integration fails or the case is sensitive.

Channels where people already are

A collaborator is published on the channels each audience already uses: a web chat widget, WhatsApp Business, Microsoft Teams for internal teams, and voice or a digital avatar when the experience calls for it. Channel connectors are configured once per organization and then assigned to each collaborator, so the same tutor can answer on the campus website and on the institution’s WhatsApp line with the same knowledge and the same rules.

What's next

Humanika is in beta and running pilots, focused first on education and then on corporate training and HR. The next steps are reusable implementation playbooks per collaborator type, more governed capabilities on the integrations above, and published results once pilots produce measurable outcomes. We will not publish numbers before then.

Learn more at humanika.co, read how we approach AI agent development and applied AI for businesses, or talk to our team about a governed AI agent for your organization.

The challenge

Universities and companies in Latin America want AI that answers with their own information and acts inside their own systems, but most tools on the market are either generic chatbots with scripted flows or open-ended assistants that cannot see institutional data, cannot execute governed actions and leave no audit trail. Each new use case ended up as a one-off build. The challenge was to turn knowledge, systems and channels into operational capacity in a way that is repeatable, secure and supervised by humans.

The solution

150 Por Ciento created Humanika, its own applied AI initiative. Humanika designs, integrates and operates digital collaborators: AI agents with a single job, such as an academic tutor, an onboarding assistant or a help desk, connected to the organization's documents, systems, processes, permissions, channels and human escalation. A control plane, Humanika Core, configures each collaborator in guided steps: client, type, knowledge, integrations, model, channels and review. The model interprets intent and requests actions; the platform validates permissions, executes those actions, records them in an audit log and hands the conversation to a person when it should. Knowledge is retrieved with RAG on PostgreSQL and pgvector, and the model layer is provider-agnostic.

4
Deployment channels
5
Solution lines
5+
Business integrations

Want to build something like this? Explore our services →

Next project

Chevignon x Formula 1: 3D car racing game
Chevignon x Formula 1: 3D car racing game

Read our Privacy Policy to learn how we protect your data and respect your privacy.