HAVOOP

01LAB

The research side of Havoop.

Havoop's research ground

Havoop is a studio. But it's also a working ground on the open problems of AI deployment in SMBs — areas neither academic research nor consumer industry are tackling.

02OPEN PROBLEMS

What we're working on.

Open problems in AI deployment for SMBs

  1. 01

    Make an AI agent hold in a heterogeneous stack without migrating everything

    The typical European SMB runs Qonto + Pennylane + HubSpot + Gmail + Notion + a vertical ERP. How do we connect an agent to that without breaking daily operations? Consumer integrations (Zapier, Make) break at scale. Enterprise solutions (MuleSoft, Boomi) are oversized. We're working on the right middle ground.

  2. 02

    Prove an AI agent's ROI in 90 days without observation bias

    The hardest metric in operational AI: real ROI, net of enthusiasm effects. We're experimenting with pre/post deployment measurement protocols that isolate the gain attributable to the agent from the gain attributable to concurrent management attention.

  3. 03

    Calibrate human-in-the-loop on non-critical workflows without prohibitive operating cost

    Classic HITL puts a human on every action. At small scale, that kills ROI. We're working on adaptive HITL architectures where the human-trigger threshold self-calibrates based on model confidence and historical error rates.

  4. 04

    Evaluate open-source vs proprietary models in real SMB conditions

    Public benchmarks (MMLU, HumanEval) say nothing about Mistral vs Claude on a French accounting workflow. We're building task-specific benchmarks for SMB use cases — accounting, customer relations, admin processing — and we publish the results.

03COMMITMENTS

What we contribute back.

Commitments and transparency

OPEN SOURCE

Whatever we build that can serve beyond our own clients gets open-sourced. First repo planned for late 2026: a generic Qonto → Brevo connector with HITL validation.

PUBLICATIONS

One in-depth article per month on the blog. Not disguised marketing: actual mission write-ups, concrete numbers, mistakes we made and corrected.

TRANSPARENCY

Public pricing. Explicit privacy policy (named subprocessors, hosting locations specified). No client data used to train models.

ECOSYSTEM

Active participation in the French AI ecosystem: France Digitale, La French Tech, lessons published on LinkedIn and at sector conferences.

04STACK

What we work with.

Our technical stack

We pick tools based on need, not the reverse. Systematic preference for European models (Mistral) when confidentiality requires it, Claude when complex reasoning is critical, GPT when the ecosystem is decisive. Open-source (Llama, Mixtral) on sovereign infrastructure for genuinely sensitive cases.

  • Claude · GPT · Mistral
  • Open-source (Llama, Mixtral)
  • Python · TypeScript
  • Vector DBs (pgvector, Qdrant)
  • Supabase · Vercel · Sanity
  • EU hosting (Scaleway, OVH)

0510-YEAR VISION

Why we do this.

Our 10-year bet

We start with French and European SMBs because that's the most under-served segment of the AI ecosystem today. Too small to interest €200k consultancies. Too specific for generic consumer solutions. Stuck in between.

The real bet is to build the operational AI infrastructure for the 25 million European businesses that can't afford a consultancy nor hire a Head of Data. Ten years to become the default AI agent for European SMBs — not a tool, an infrastructure.

Everything we build in services today — diagnostic, deployment, production — funds the research that will make the AI agent as standard as a Qonto account for a European SMB by 2035.