AI architect · Product builder · Founder

Chengheri
Bao.

I build systems that solve real problems—from enterprise AI infrastructure to software I need in my own life as a founder.

LinkedIn See selected work

Portrait of Chengheri Bao — PhD Physics · 8+ years AI · Paris

About

I’m an AI architect, engineer and product builder. For the past 8+ years, I’ve worked on AI, data and software systems inside large organizations—from predictive models and data products to production-grade AI architectures.

My background is in engineering and physics (PhD), but most of my career has been spent turning technical ideas into systems that actually work in the real world.

Today, I work in two directions: helping teams bring AI systems into production and building my own software products from problems I encounter myself.

Engineer → Architect → Builder → Founder.

Selected experience

From research to production.

Systems built across telecoms, banking, insurance and gaming—with architecture, evaluation and production engineering at the center.

Case study / 0104 2025 – 04 2026

Agentic AI · Enterprise architecture

Agentic System

Bouygues Telecom

Make architecture knowledge easier to find and use.

Outcome

Cut technical-report drafting time by 50% while giving architects a single assistant for use-case discovery, reference mapping and grounded technical answers.

01 / Challenge

Enterprise architects need to locate relevant use cases, connect requests to the company’s reference framework and answer technical questions across dispersed documentation.

02 / System

The system brings together three capabilities: finding relevant use cases, aligning requests with enterprise reference frameworks and answering technical questions from validated internal documentation. It presents the supporting context for each response and gives experts a structured way to review and maintain the knowledge provided.

Signal to system

01
Architecture question
02
Intent routing
03
Search & mapping
04
Grounded answer

Technology

PythonLangChainLangGraphFastAPIMongoDBNeo4jWeaviateTransformersHugging FaceRAGASDoclingStreamlitspaCyAzure

AWSMLflowOpenShiftKubernetesHelmKustomizeDockerGitLab CI/CDLocust

My role

Tech Lead AI: architecture, agent orchestration, backend and interface development, deployment, evaluation and knowledge transfer.

Case study / 0203 2022 – 11 2024

Anomaly detection · Responsible gambling

Anomaly Detection

FDJ · Française des Jeux

Surface unusual activity for teams to investigate.

Outcome

Increased the detection rate 10× and shortened detection time from days to hours—shifting teams from reactive investigation to proactive monitoring.

01 / Challenge

Responsible-gambling teams need to identify unusual patterns in retail activity from large volumes of real-time transactional data, with limited labeled examples.

02 / System

The system processes real-time transaction activity, consolidates it into shop-level risk indicators and generates prioritized alerts with supporting explanations. Analysts receive recurring scoring and reporting views that let them examine unusual behavior, compare cases and document their investigations.

Signal to system

01
Real-time transactions
02
Rolling features
03
Anomaly scoring
04
Explain & report

Technology

PythonApache SparkHiveHDFSTensorFlowScikit-learnScikit-TDAPandasNumPyMatplotlibPlotlySciPySHAPMLflowKubernetesAirflowDockerGitLab CI/CDElasticsearchKibanaPower BIGit

My role

Lead data science and production delivery: risk modeling, data pipelines, experiment tracking, monitoring and stakeholder reporting.

Case study / 0305 2021 – 01 2022

Predictive modeling · Explainability

Risk Model

Allianz

Improve credit assessment while keeping decisions understandable.

Outcome

Made 100% of model decisions explainable while improving predictive performance by 15%.

01 / Challenge

Credit assessment needs predictive strength and a model that risk teams can understand, review and document.

02 / System

The system turns applicant and business information into a credit-risk assessment, then presents the factors behind each score for review by risk teams. It also supports validation across time periods, stability checks and the documentation required for governance and regulatory review.

Signal to system

01
Business data
02
Feature engineering
03
Hybrid model
04
Explainable assessment

Technology

PythonLogistic regressionXGBoostDecision treesFuzzy logicScikit-learnPandasNumPyMatplotlibSHAPKedroMatlab

AWSGit

My role

Lead data science: assessment of the expert system, hybrid model development, feature engineering, explainability and technical documentation.

Case study / 0402 2020 – 04 2021

Conversational AI · Semantic matching

Conversational AI

Crédit Agricole

Connect banking questions with the right knowledge.

Outcome

Reduced reliance on human call support by 40%, deployed across 39 regional banks and delivered responses up to 8× faster.

01 / Challenge

Banking advisors need precise answers across specialist business domains, even when their wording differs from the language in the knowledge base.

02 / System

The system receives an advisor’s question, identifies its business domain and intent, and returns the most relevant approved answer. It supports specialized vocabulary across eight banking domains and includes a structured process for reviewing and updating the available knowledge.

Signal to system

01
Advisor’s question
02
Language & entities
03
Semantic matching
04
Relevant knowledge

Technology

PythonBERTFastTextPyTorchScikit-learnGensimNLTKspaCyTF-IDFNERTesseract OCRPandasNumPyMatplotlibElasticsearchMongoDBJenkinsKibanaGit

My role

Lead AI development and optimization: language models, intent matching, semantic retrieval and production integration.

Now building

Products born from lived problems.

The same discipline runs through every product: notice a problem, build the tool I wish existed, use it extensively, and learn before offering it to others.

01 / Content