Skip to content

Carnival DevOps · Bangkok · Worldwide

Systems that keep turning.

We design, run and secure the infrastructure behind your business, then give your AI a memory of it. Twenty years on call, four products, one obsession: uptime.

--:--:--Bangkok
--:--:--Paris
--:--:--Stockholm
Scroll

Products

Four products.
One obsession.

Each one started as a tool that a client or an agent needed. Ferris and Recall run on hardware you control. Kermes works in the field, with or without a network. Delectable keeps a personal library on your own disk.

FerrisManaged monitoring · Endpoint console

Self-hosted · White-label

Monitoring your clients can actually read.

Your client should never be the one who tells you a server is down. Ferris watches every machine, service and certificate at a site, warns whoever is on call, and proves at the end of the month what it saved them. Underneath: metrics and alerting, service and TLS probes, dashboards, endpoint inventory and patch status, remote scripts, floor plans, maintenance windows and a monthly expert report — one stack, one configuration file, validated before it touches anything that is running.

  • 24/7 alerting to Telegram, e-mail or webhook
  • HTTP, TCP, SMTP, IMAP and TLS-expiry probes
  • Linux, macOS and Windows agents
  • Auto-healing hooks triggered by alerts
  • Floor plans, switches and cabling maps
  • Monthly PDF report, English or French
  • Zero-touch enrollment of new machines
  • Optional SIEM and intrusion detection
20+Alert rules out of the box
1Config file to deploy
3Agent platforms
24/7Watch, with a human behind it

RecallMemory layer for AI agents

Self-hosted · MCP

Memory that understands, not search that guesses.

An assistant that starts from zero every morning costs you the same explanation every morning — and cheerfully proposes again the fix you rejected last month. Recall gives coding agents and assistants a memory of what was decided, and why. Every query runs through hybrid BM25 + semantic search, cross-encoder reranking, recency scoring and knowledge-graph boosting, so the agent gets the right memory rather than a similar one. Ships as an MCP server for Claude Code and a REST API for everything else, and runs entirely on your hardware.

Concretely — a developer and their agent spend the first ten minutes of every session re-establishing what the system is, which database is authoritative and what was already tried in August. With Recall the agent asks once, and the decision from August comes back with the reason it was taken.

  • Hybrid search with cross-encoder reranking
  • Knowledge graph with entity extraction
  • Conflict detection between memories
  • Document distillation, page by page
  • MCP server + REST API
  • Multi-tenant with row-level security
  • GPU or CPU, AMD or NVIDIA
  • Your data never leaves your network
~60 msSearch latency
4Stage retrieval pipeline
31MCP tools exposed
100%Self-hosted

KermesInspections & audits · Offline-first

White-label · iOS · Android · Web

Inspections that work offline.

A paper checklist is filled in once and typed up twice: in the evening, into a spreadsheet, from photos nobody renamed — and when a tenant disputes a scratch, none of it proves anything. Kermes is a white-label app for inspections and audits. Inspectors fill forms on iOS or Android with no network, with photos, video and comments. Managers review, resolve conflicts, lock the inspection and generate the report in the web dashboard, where every answer keeps its author, its time and its history.

  • Offline answers, photos, video and comments
  • Form builder and Excel import
  • One Sync button: push, then pull
  • Conflicts kept for a manager to choose
  • Annotations and PDF reports
  • Locked inspections, admin-only unlock
  • Change history: who, when, old and new value
  • Your name and logo on every app
8Field types, from yes/no to photo, video and repeating tables
100 %Of an inspection works offline
3Apps: iOS, Android and web
3Roles: inspector, manager, admin

DelectablePersonal library · Desktop

Early access · Linux · Windows · macOS
Delectable: shelves of film posters and cloth-bound book covers

Your Delicious Library, back on the shelf.

Delicious Library, the Mac app that put books, films, music and games on virtual shelves, stopped at the end of 2024. Delectable opens its library file — items, covers, notes, shelves, ratings — and keeps your original untouched. Then it adds what a shelf app never had: search by title in any language, scores, cast and where to watch.

  • Imports a Delicious Library 3 file, twice-safe
  • Add by title or by barcode, or from a photo
  • IMDb, Rotten Tomatoes and Metacritic scores
  • Summary, cast and where to watch
  • Complete items without replacing what you typed
  • Loans, notes, ratings, themes
  • French and English
  • A folder on your disk: no account, no cloud

What we deliver

From the rack to the retrieval layer.

We work where infrastructure meets AI. The same team that segments your network also builds the knowledge graph your agents query.

01

RAG & Knowledge Graphs

Retrieval pipelines that behave like memory: hybrid search, cross-encoder reranking, entity extraction, graph databases, MCP integration for coding agents.

pgvectorNeo4jbge-m3MCP
02

Infrastructure Architecture

Proxmox and bare-metal virtualisation, hybrid cloud, high-availability clusters, storage, zero-downtime migrations, network segmentation and firewalls.

ProxmoxLXCOPNsenseZFSCloudflare
03

AI Systems & LLM Ops

Self-hosted LLM serving on AMD and NVIDIA GPUs, embedding and reranking services, agent orchestration, evaluation and cost control.

vLLMllama.cppROCmCUDAFastAPI
04

Managed Monitoring & SOC

Ferris deployments with 24/7 alerting, service and certificate probes, endpoint inventory and patching, optional SIEM and intrusion detection.

FerrisAlertingEndpointsSIEM
05

Odoo ERP

Full-lifecycle Odoo 17 and 18 deployments: PostgreSQL tuning, custom modules, integrations, legacy migrations, Google Workspace.

OdooPostgreSQLDocker
06

DevOps & Automation

CI/CD from GitHub, Docker orchestration, Terraform and Ansible, workflow automation with n8n, security hardening and audits.

GitHub ActionsTerraformAnsiblen8n
07

Websites & 3D Experiences

Fast, hand-built sites with real-time 3D when it earns its place: Blender-made models, WebGL scenes, motion that runs at 60 fps, clean SEO, deployed on the edge. This page is one.

Blenderthree.jsWebGLCloudflare PagesSEO

Our approach

The work beneath the surface.

Five steps, always in this order. Skipping one is how systems end up needing attention at 3 a.m.

01

Observe

We start with what actually runs: the inventory, the dependencies, the failure mode nobody has written down yet.

02

Design

The smallest architecture that survives a bad day. Boring components, explicit boundaries, one source of truth.

03

Deploy

Everything ships from Git and is validated before it touches production. Rollback is a decision, not an adventure.

04

Automate

Alerts that page the right person once. Reports that write themselves. Agents that remember the last incident.

05

Celebrate

Quiet dashboards, clients who sleep, and a system you could hand to a junior technician with the runbook.

Why Carnival

An independent perspective.

No reseller margins, no platform to push. We recommend what we would run ourselves, because we usually already do.

On call since 1998

Enterprise IT across Ireland, Sweden, Hong Kong and Thailand. We have been the person paged at 3 a.m., which is why we design so that nobody is.

Sovereign by default

Self-hosted first. Your metrics, your documents and your AI memory stay on hardware you control. PDPA and GDPR are design inputs, not afterthoughts.

Built to hand over

Runbooks written for the junior technician, monthly reports your management can read, no vendor lock-in and no secret sauce you cannot audit.

A good system is like a ferris wheel: it turns all night, and nobody thinks about the motor.

Frequently asked

Questions we get a lot.

What exactly does Carnival DevOps do? +
We architect, run and secure infrastructure — on-premises, Proxmox, hybrid cloud — and we build the AI layer on top of it: retrieval, knowledge graphs, agents with memory. Four products came out of that work: Ferris, so a site tells you it is in trouble before the client does; Recall, so an AI assistant remembers what was decided; Kermes, so an inspection done with no signal still produces a report you can defend; Delectable, so a library catalogued for years in Delicious Library is not lost when that app is.
Is Ferris a SaaS? +
No. Ferris is deployed at your site, or on a host we operate for you, white-labelled, with your data staying on your network. You pay for the retainer and the expertise, not per endpoint.
Can Recall run fully offline? +
Yes. Recall runs on your hardware, GPU optional. Embeddings, reranking and the knowledge graph never leave your network. It exposes an MCP server for Claude Code and a REST API for everything else.
Where are you based, and who do you work with? +
Bangkok, Thailand (UTC+7), working with clients in Southeast Asia and Europe. French, English, Swedish, Thai and Spanish spoken.
Do you take small projects? +
Yes, if they are well defined: an audit, a migration, a monitoring rollout, a RAG prototype. We say no to projects we do not understand.
How do we start? +
Write to hello@carnival-devops.com, or reach out on LinkedIn, with a short, non-confidential description of the problem. We reply with questions, then a written proposal. No sensitive documents in a first message, please.

Let's talk

Let's build something that keeps turning.

Tell us about the system that worries you, or the one you wish existed. A clear reason for reaching out is enough to begin.

hello@carnival-devops.comBangkok · UTC+7EN · FR · SV · TH · ESReplies within one business day