01
An agent that runs on your premises
It reads your business tools.It answers, sorts, follows up.On your server.Nothing leaves.
A sovereign, autonomous agent, on open-weight models you can upgrade.
An engineering studio, founded by former colleagues specialised in machine learning and big data.
01The math
You pay for AI twice. We stop the meter.
Once for the subscriptions and token bills that arrive every month. Once for the hours your teams still spend chasing quotes and rebuilding the same reports.
Run the numbers with your own figures
Three sliders, nothing sent anywhere: everything is computed in your browser.
What you pay today
- per month
- €925
- per year
- €11,100
- over three years
- €33,300
Meter since January 1st
€0.00What these subscriptions have already cost this year, to the second, with the values above.
What it looks like on your side
Five ways to run the same thing. The one we would recommend for your figures is marked.
Our part is one flat fee, paid while we build and install. Once the system is delivered, it is 0 EUR per month, in every case. You only come back to us for a version upgrade or a change of model.
Wherever it runs
Several people at once, immediate responses. The server stays on your premises and your data never leaves them.
- With Wikolabs, after delivery
- 0 EUR per month
- What it takes
- 1 machine
The server, the workstation or the hosting are bought separately, from your own supplier. We take nothing on them and nothing is paid to us afterwards.
What you stop paying : €925 per month€11,100 in the first year
Our installation fee is quoted after an audit of your estate.
Send us this calculation, we reply with an installation ballpark within 48 hours.
02In short
We install machines that work on their own, on your premises.
Six things, and nothing else. Each one is detailed in the next section, with its spec sheet line by line.
02
Data that speaks
Your six files become one number.Updated overnight.Readable at eight.No per-seat licence.
03
A prototype that holds
It worked in a demo.We make it robust.Balancer, replicas, restored backups.Not a rewrite.
04
A box that decides on site
It reads cameras and meters.It decides without the cloud.The network drops.The line keeps running.
05
An audit that ends in a decision
A week inside your processes.Use cases ranked.Budget per batch.Not one more report.
06
What it actually costs
One payment, priced after the audit.Then tens of euros a month.The code and models stay yours.You can leave.
This is not one more subscription. It is an installation.
03Our offers
Five ways to work with us.
From a one-week audit to an agent installed on your premises. Every offer comes with its spec sheet.
01Timeline 3 to 8 weeks
Sovereign agent on open-weight models
Plugged into your tools, fine-tuned on your data, installed on your premises.
- Connectors for ERP, CRM, databases, messaging
- Open models on your server, offline where you want
- Fine-tuning on your texts, audio and images
- Model weights and training recipe handed over
Typical installation sheetSovereign agent
- Text generation
- medium open model, 7 to 14 B
- Semantic search
- multilingual embeddings
- Vision and OCR
- vision-language model
- Voice
- transcription, synthesis, diarization
- Vector store
- built into your database
- Hardware
- 32 GB and a quantised model, or a dedicated GPU
- Outbound data
- none
- Monthly cost
- close to zero
Configuration adjusted to your hardware and volumes before signature, never after.
04The catalog
Forty-eight agents and systems, sorted by function.
Each line is a production-tested system, built on specialized open-source models and pluggable into your stack. The detailed architecture lives on each offer's page.
05The method
Six steps, from discovery call to production.
Daily check-in, weekly review, a demo at every sprint. You validate each deliverable before we move to the next.
- 01Discovery call (30 min)
A call to understand your context, challenges, and goals. We identify together where AI can create the most value for you.
- 02Scoping & proposal
We structure use cases, user stories, and business rules with you. You receive a clear proposal with scope, timeline, and delivery milestones.
- 03Kick-off & access
Team introduction, collaboration tool setup, and access provisioning. The project starts in a structured and secure environment.
- 04Sprints & bi-weekly demos
Daily standup, 1-2 week sprints, end-of-sprint demo. You validate every deliverable before moving to the next phase.
- 05Testing, QA & production launch
Unit tests, UAT (User Acceptance Testing), bug fixes, and production deployment on your infrastructure or ours.
- 06Support, optimization & roadmap
Post-delivery monitoring, performance tracking, new features, and scaling. Weekly review to steer the roadmap together.
1 of 6
06The team
Six roles, one team, the whole lifecycle.
From the agent engineer to the MLOps engineer: the profiles that carry your project from the first data pipeline to production monitoring.
Agentic AI engineer
Designs and deploys autonomous AI agents with memory, skills and tools: sales agents, support agents, market intelligence agents.
Data engineer and architect
Builds real-time data pipelines, warehouses (PostgreSQL, ClickHouse, BigQuery, Snowflake) and CDC architectures that power your agents and dashboards.
ML and LLM specialist
Fine-tunes, trains and deploys custom models (RAG, RLHF, multimodal embeddings, speech, vision) for your domain, your vocabulary and your accents.
Edge computing integration
Runs models next to the field: IP cameras, sensors, Raspberry Pi, NVIDIA Jetson, industrial boxes, with local inference, offline operation and deferred sync.
Automation specialist
Orchestrates your workflows with n8n, Zapier, Make and LangGraph pipelines to connect your tools and remove repetitive tasks.
MLOps and cloud engineer
Deploys, monitors and maintains your models in production on GCP, AWS or Azure, with CI/CD pipelines and full observability.
07The team
Different paths, one conviction: data first, decisions next.
Wikolabs is a team of engineers. Each of them writes here what they have understood about their craft, and the path that led there.
COST KILLERGuillaume Rakotonjanahary Tsantaniaina Amboarasoa
Founder and Principal Architect
MSc Big Data and AI, ESTIA France
The path
- 01Full-stack developer and data engineer Web and mobile applications, data pipelines, first warehouses.
- 02Data scientist and machine learning engineer Forecasting, classification and recommendation models shipped to production.
- 03AI engineer and tech lead Agents, RAG, vision and voice; technical leadership of bilingual teams.
- TodayFounder of Wikolabs Free companies from recurring subscriptions to external services: an autonomous, sovereign intelligence system, a smart and predictive data infrastructure, a monthly cost cut to the minimum.
Most CEOs I meet are paying for AI twice. Once for the subscriptions and token bills that arrive every month, and once for the hours their teams still spend chasing quotes, answering the same questions, and rebuilding the same report in a spreadsheet. The AI is running. The work is still manual.
The systems that move a business are not chat assistants. They are agents that follow up on every quote you sent last month, answer the phone, read an invoice into your ERP, flag a defect before it ships, and tell you on Monday what changed in your numbers. Those systems do not have to run on a metered API.
A generalist model is extraordinary at open-ended reasoning, and you should keep one for that. But for the work an agent actually does, classifying a message, extracting a field, scoring a lead, retrieving the right document, a specialised open-source model does the job as well or better, on hardware you already own. The bill stops. The work does not.
That is what I build.
I start with the data layer, because that is where AI projects fail far more often than at the model: ingestion pipelines, warehouses, semantic layers and BI dashboards that give you one honest view of your business. Then the agents, websites and agentic apps on top.
Computer vision shows what ownership buys you: defect detection on a line, PPE and safety compliance, shelf and stock counting, OCR on invoices and delivery notes, site access control. High volume, low value per call. Metered pricing makes it absurd. Owned hardware makes it routine.
Voice is the honest exception. Natural conversation needs sub-second latency, reliable voice activity detection and clean turn-endpoint detection, and the best of that still comes from hosted providers. So I use them where they earn it and engineer around the cost: local models for understanding and retrieval, hosted only for the speech layer, tight context control so a long call does not become a large invoice. You pay the provider and the carrier directly, never a marked-up platform fee, and I say so before you sign.
You pay for the build once. You own the stack, your data stays in your infrastructure, and the only reason to come back is when a genuinely better open-source model ships.
If you carry a monthly AI bill and cannot point to what it replaced, that is worth a conversation.
08The hub
Madagascar, the hub for sovereign AI integration.
Thanks to local expertise trained on open models, data and on-site deployment, Antananarivo installs sovereign systems for Europe, Canada, the French Caribbean and Guiana, the Indian Ocean, Africa and India. UTC+3, delivery in French and English.
- Engineers trained and based in Antananarivo
- Installed on your premises, on your server or your cloud, wherever it is
- Meetings in your time zone, documentation in your language
09Our references
Our partners and our references.
Partners
You build a business solution and want sovereign agents inside it? Let's talk.
The tools we operate
dbt
Twilio10The journal
What we write for executives.
Analyses on agentic AI, automation and data, without the jargon.
