AI & SaaS · Our Own Product 2023-2026

Nexo (producto propio)

Nexo is Kiwop's SaaS platform that brings together CRM, projects, invoicing and time tracking with AI agents that query data and execute tasks with permissions, human approval and traceability.

3 years (ongoing) 6 specialists 2023-2026
Nexo — The platform
Services
  • AI Agents
  • Laravel Development
  • React Development
  • API Integrations
  • UX/UI Design
  • DevOps
Related services

01The challenge

The challenge

Companies use dozens of disconnected tools: CRM, project management, time tracking, invoicing... The result is data duplication, manual processes and zero real visibility. Nexo was born to solve this at home first: we wanted to run Kiwop with a single system that learned from our own data. We're our own first client, and that's the point: it's the same AI brain we build for whoever hires us.

02The solution

The solution

We developed Nexo as a multi-tenant platform with modular architecture: time tracking with legal compliance (4-year retention), project and task management, CRM with lead capture by email and invoicing synced with Holded. On top, a cross-cutting AI layer: an assistant that answers with the company's data through audited tools, and 27 proactive agents that monitor time budgets, invoicing, incoming mail and SEO, generate reports and propose actions. Every action with consequences requires human approval and is logged. Live integrations with Holded, Apollo, Slack and the Google APIs, plus importers for HubSpot, ClickUp and Asana. Laravel + React + Inertia.js stack on PostgreSQL with pgvector.

Tech stack

How it's built

Front end

Engine

AI

  • pgvector Vector search
  • Anthropic API Claude models via API
  • Claude Opus 4.8 Strategic LLM
  • Claude Sonnet 5 Workhorse LLM

Data

  • PostgreSQL Database
  • Redis In-memory cache

03Architecture

How a request flows

  1. An event (a cron job, an incoming email, a user question) triggers an agent or the assistant

  2. The system retrieves only the authorized context: every query is filtered by workspace and by the user's or agent's permissions

  3. The model reasons and picks a tool from a catalog of 37: searching tasks, hours, leads, profitability, internal knowledge or data connectors

  4. Read-only tools run directly; the ones that mutate something (write SQL, files, shell) require the user's explicit confirmation

  5. Agent action proposals are born in a pending state and only run once a human approves them

  6. Every call to the model and to each tool is logged: tokens, cost, duration and who approved what

What is AI and what is deterministic

AI components

The models (Claude Opus and Claude Sonnet, depending on each task's difficulty) interpret instructions, reason about the company's context, choose tools, draft reports and drafts, and propose actions. Routine agents run on the cheaper model and strategic ones on the most capable: model tiering is part of the design.

Deterministic components

Invoicing sync (Holded), time tracking and its 4-year legal retention, isolation between workspaces, role- and module-based permissions and business validations all run through conventional code. We never delegate to the LLM any operation that requires a deterministic outcome: the model proposes, the code decides what it can touch.

What the system does not do

  • No agent accesses data outside its workspace: isolation is enforced on the server (scopes, middleware and dedicated tests), not in the prompt
  • Every action proposal from an agent is born in a pending state and requires human approval before it runs
  • Tools that mutate data (write SQL, files, shell) require explicit confirmation and are restricted to authorized users
  • Time tracking and legal obligations run through deterministic logic, never through the model's inference
  • Every AI call is metered in an immutable log (tokens, cost, model) with a monthly budget per workspace that cuts off spending if it's exceeded

04Measured impact

Results

  • 10+ Live integrations
  • 90% Automated time entries
  • 100% Time tracking compliance
  • 27 Active AI agents

Verifiable data

Metrics in detail

Months in production (platform / agent layer)
36 / 5
Projects managed
166
Tasks managed
2 022
Time entries (clock-ins)
2 478
Hours billed to client projects
2 983 h
Time entries created via automated channels (API, task closure, git)
90%
Active AI agents
27
Agent executions logged (since Feb 2026)
6 760
Model-to-tool calls logged
466
Tokens processed by the AI layer
142 M

Methodology: Aggregated data extracted via a read-only query against Nexo's production database on July 10, 2026. The platform has been in production since July 2023 (it started as time tracking) and the AI agent layer since February 2026. These are platform-wide totals; they expose no client data.

05The process

The process

  1. Discovery

    Operational workflow analysis and pain point identification

  2. Architecture

    Multi-tenant design and integrations model

  3. Core

    Time tracking and project management

  4. Integrations

    Holded, Apollo, Slack and Google APIs live; importers for HubSpot, ClickUp and Asana

  5. SaaS

    Onboarding, billing and admin panel

  6. AI Layer

    Assistant with audited tools and proactive agents with human approval: reports, monitoring and proposals

Your next project?

Let's talk results

Tell us your challenge. We'll respond within 24h.

Start a conversation
  • NDA available
  • Response <24h
  • Phased proposal
Next project Kiwop (activo propio) Technical SEO & Citability Architecture