The AI practice of Cogniflamme Labs

Efficient, private AI systems for business workflows

We design and deploy lightweight AI — knowledge assistants, workflow automation, multi-agent business systems and fine-tuned small models — that run where your data lives. No hype, no vendor lock-in. Just systems that work.

Explore ServicesStart a Project
flammai — session

$ flammai ignite --audit ./your-workflows

scanning workflows .......... 14 found

ranking by roi .............. 3 candidates

building proof of concept ... 14 days

✓ smallest system that works — private, cited, evaluated

$ ▊

scroll
Private & Local AI✦RAG Knowledge Systems✦Small Fine-Tuned Models✦Workflow Automation✦Multi-Agent Business OS✦GDPR-Friendly by Design✦Open Standards✦Evaluation-Driven✦Production over Demos✦

// Services

One principle: the smallest system that works

Most businesses don't need frontier-scale AI. They need reliable, efficient, private systems that automate real workflows. Seven ways we build them:

01

Ignite

AI Discovery Sprint

Details
02

Pulse

Private Knowledge Assistants

Details
03

Flow

AI Workflow Automation

Details
04

Beacon

Business Operating Systems

Details
05

Forge

Custom Model Adaptation

Details
06

Hearth

Private & Local AI Deployment

Details
07

Hefesto

Open-Source R&D

In developmentDetails

Not sure which one fits?

Start with an Ignite sprint
01

Ignite

AI Discovery Sprint

Every engagement starts with evidence. In two weeks we go from “we should do something with AI” to a validated use case with a working proof of concept — built on your data, with measured results.

Book a sprint

What you get

  • A workflow and data audit: where AI actually pays off in your operation
  • A use-case map ranked by ROI, feasibility and risk
  • A working proof of concept on your own documents or data
  • An architecture recommendation with realistic cost projections

Format

Fixed scope2 weeksFixed priceGo / no-go clarity
02

Pulse

Private Knowledge Assistants

Your company's knowledge is scattered across PDFs, wikis, tickets and inboxes. Pulse turns it into an assistant that answers with sources — and keeps every document inside your perimeter.

Discuss your documents

What you get

  • Ingestion pipelines for PDFs, office files, wikis, tickets and archives
  • Hybrid retrieval (vector + keyword) with reranking for precise answers
  • Source citations on every answer — no unverifiable claims
  • Access control, evaluation sets and a feedback loop that improves over time

Stack

pgvectorQdrantHybrid searchRerankingOpen embeddings
03

Flow

AI Workflow Automation

The profitable side of AI is boring: invoices extracted, emails triaged, contracts summarized, data entered. Flow builds deterministic pipelines with human approval where it matters — predictable automation, not autonomous chaos.

Automate a workflow

What you get

  • Structured extraction from invoices, contracts, CVs and operational documents
  • Email and ticket triage: classify, route, draft responses for human review
  • Multi-step pipelines with validation rules and human-approval checkpoints
  • Schema-validated outputs that plug into your CRM, ERP or database

Stack

LangGraphPydanticAIOCRJSON SchemaTemporal
04

Beacon

Business Operating Systems

Beyond answering questions: a team of specialised agents that runs your recurring operations — follow-ups, back-office, reporting, support — relaying work like a chain of signal fires. Each agent has a narrow role, scoped permissions and an audit trail; people approve what matters.

Design your agent team

What you get

  • An agent team mapped to your real operations: roles, responsibilities and hand-offs
  • Tool access through MCP and your APIs — scoped permissions, sandboxed execution
  • Shared memory and context across agents, grounded in your own knowledge base
  • Approval gates, full audit logs and a kill switch — autonomy you can supervise

Stack

OpenClawHermes AgentRig (Rust)MCPLocal models
05

Forge

Custom Model Adaptation

When a workflow repeats thousands of times, a small model tuned to your domain beats a frontier API on cost, latency and privacy. Forge is where we shape it — trained on your patterns, evaluated against your reality.

Forge your model

What you get

  • LoRA/QLoRA fine-tuning of open models (Mistral, Llama, Qwen, Gemma)
  • Dataset curation and synthetic data generation from your examples
  • Quantized builds that run on modest hardware — even CPU
  • Evaluation reports: your model vs. frontier APIs on your actual tasks

Stack

UnslothAxolotlPEFTQLoRAllama.cpp / GGUF
06

Hearth

Private & Local AI Deployment

Some data can never leave the building — and it shouldn't have to. Hearth puts the whole stack on your infrastructure: models, retrieval, monitoring. GDPR-friendly by design, maintained like production software.

Bring AI in-house

What you get

  • On-premise or private-cloud inference: Ollama, vLLM, llama.cpp
  • Full-stack deployment: models, vector store, APIs, authentication
  • Observability: traces, cost, latency and quality dashboards
  • Managed maintenance: updates, evaluation, capacity planning

Stack

OllamavLLMDockerGPU / CPU servingObservability
07

Hefesto

In development

Open-Source R&D

Our research bet, named after the god of the forge: a Rust runtime where AI capabilities are forged from your APIs, databases and tools as reusable skills — composed into workflows and routed to the smallest capable model.

Follow on GitHub

Direction

  • Skills as the first-class abstraction — not agents, not giant prompts
  • Model-agnostic routing: the right model per task, small first
  • Open standards: MCP, OpenAPI, JSON Schema — local-first execution
  • Open core: a self-hostable runtime and a community skill registry

Stack

RustMCPOpen sourceLocal-first

// Principles

How we think about AI

  1. 01

    Efficiency over scale

    The goal is never the biggest model. It's the smallest system that does the job.

  2. 02

    Privacy first

    Your data stays where it belongs. Local inference and GDPR-friendly architectures by default.

  3. 03

    Composition over monoliths

    Reusable capabilities, not giant prompts. Systems you can inspect, test and extend.

  4. 04

    Open by default

    Open-weight models, open standards, no vendor lock-in. You own what we build for you.

  5. 05

    Production over demos

    Evaluation, observability and maintenance built in. Demos are easy — reliability is the product.

// Process

From use case to production

01

Discovery

Map your workflows, data and constraints to find real ROI.

02

Prototype

A working proof of concept within weeks, tested on your data.

03

Evaluate

Measure accuracy, latency and cost before scaling anything.

04

Deploy

Ship to your infrastructure with observability built in.

05

Evolve

Monitor, maintain and improve as your data and needs change.

// About

The AI practice of Cogniflamme Labs

FlammAI is where Cogniflamme Labs concentrates its AI engineering: an independent, senior-level practice based in France, working with clients worldwide. We combine deep infrastructure experience (DevOps, security, cloud) with applied AI engineering — which is exactly what it takes to move AI out of the demo phase and into production.

// Contact

Start a conversation

Describe your project and we will get back to you. Or reach us directly at info@cogniflamme.com