Research and evaluation
Turn an uncertain objective into a measurable technical problem. Establish the baseline, build the evaluation set, locate the failure and test competing approaches.
- Evaluation
- NLP
- Retrieval
- Model adaptation
CORNICHON · Paris / Worldwide
Applied AI researchand engineering.
We work across models, evaluation, data, infrastructure and product when the baseline has plateaued, the architecture remains unresolved, or the system still has to be built.
Start a conversationQuestionNo established
path
path
ResearchEvaluate
and decide
and decide
LiveIn useOperating
system
system
What we do
Research and engineering across the system.
Architecture follows the evidence. Depending on quality, data access, latency, cost and governance, that may mean a tuned model on private infrastructure, a managed API, retrieval, tools, or a simpler system.
Zero-to-one engineering
Build the complete path from data and model to services, deployment and the interface people use.
- Data
- Backend
- Infrastructure
- Product
Production improvement
Measure and improve an existing system across quality, latency, cost and reliability. Add the controls needed to operate it responsibly.
- MLOps
- Feedback
- Governance
- Efficiency
Experience behind CORNICHON
Selected engagements.
Work completed by CORNICHON’s founder in client teams and consulting engagements.
SNCF Connect & Tech
Led the NLP evolution of two large customer-facing railway chatbots, including transformer migration, production retrieval, synthetic data, evaluation and direct platform integration.
- +13% F1
- 96% recall @k=2
- −74% inference cost
95%lower response latency
Sanofi
Designed an evaluated agentic workflow that reconciles clinical-trial safety narratives across hospital forms and pharmacovigilance records.
- 90% F2
- Clinical-user evaluation
- Reviewer feedback loop
4×review throughput
APICIL
Built and deployed a retrieval system for insurance brokers, covering corpus design, PDF ingestion, hybrid retrieval, evaluation and a production API.
- Business-query test set
- PDF ingestion
- Production API
91%recall @k=2
Conseil d’État
Built and integrated retrieval, reranking, summarization, OCR and PII anonymization for legal document processing, with benchmarking against GPT-4.
- Fine-tuned summarization
- PII anonymization
- Model benchmarking
Deployedlegal NLP system
Tell us what you are trying to make work.
An unresolved technical question, an underperforming system, or a new product without an established path.
Start a conversation