CORNICHONStart a conversation

AI engineering studio · Paris / Worldwide

Hard problems.Complete systems.

We solve high-stakes AI problems where off-the-shelf answers stop, then build the entire 0→1 system around the solution.

Discuss your problem
InputUnmapped
territory
CORNICHON / 01Research
+ build
LiveOutputWorking
system

Model → data → infrastructure → interface

Selected outcomes

95%less latency
74%lower inference cost
4×review throughput
201+research citations

What we do

From uncertain idea
to reliable operation.

Senior research depth and full-stack execution, without the handoffs.

01

Hard AI problems

When the obvious approaches plateau, we research, prototype and evaluate until the system works. Multi-agent systems are used when specialization and orchestration create real leverage.

  • Agents
  • RAG
  • NLP
  • Fine-tuning
02

Complete 0→1 systems

One accountable engineering team from raw idea to a usable, tested product in production.

  • Product
  • Data
  • Backend
  • UI
03

AI infrastructure

Governed, efficient foundations for real traffic, with tuned models, private deployment and feedback loops that improve the system over time.

  • Governance
  • Efficiency
  • Evals
  • MLOps

Selected work

Proof over promises.

Production systems for organizations where quality, speed and reliability matter.

01

Conversational AI · Production systems

SNCF Connect & Tech

Evolved two high-traffic customer chatbots through transformer migration, a production RAG agent and direct platform integration.

  • +13% F1
  • 96% recall @k=2
  • −74% inference cost
95%latency reduction
02

Agentic systems · Clinical operations

Sanofi

Built an eval-first agentic system that reconciles clinical-trial safety narratives across hospital and pharmacovigilance records.

  • 90% F2
  • Human feedback loop
  • Production evaluation
4×review throughput
03

Insurance AI · NLP engineering

APICIL

Built and deployed a production RAG agent helping insurance brokers find answers and present relevant offers from complex documents.

  • PDF ingestion
  • Hybrid retrieval
  • FastAPI production API
91%recall @k=2
04

Public sector · Document intelligence

Conseil d’État

Modernized legal document processing with retrieval, summarization, OCR and privacy-aware handling for a public institution.

  • RAG pipeline
  • PII anonymization
  • Model benchmarking
RAGproduction pipeline

How we work

One clear path
from 0 to 1.

Small loops, explicit measures and one team accountable across the full system.

  1. 01

    Frame

    Turn the ambiguous problem into a falsifiable target and a practical evaluation set.

  2. 02

    Prove

    Attack the riskiest assumptions first. Prototype, measure and choose the simplest viable path.

  3. 03

    Build

    Engineer the full system: model, data, infrastructure, APIs and the interface people use.

  4. 04

    Operate

    Deploy with observability, feedback loops and cost controls so performance compounds.

Built for hard problems

Research depth. Production discipline.

CORNICHON AI combines language-model expertise with the infrastructure and product engineering needed to make AI useful.

From regulated industries to high-traffic platforms, we take ownership of the full path: framing, evaluation, models, data, infrastructure and interface. We tune and host models on company infrastructure when that is the right choice. We use external APIs selectively, where their capability justifies the dependency.

Senior executionEvidence-led decisionsEnd-to-end ownership

Have a hard problem?

Let’s make it work.

Bring the challenge. We’ll bring the path from first principles to production.

Start a conversation