|
📍 Saint Jean Bonnefonds (42650)

TANGUY PAUWELS // ML ENGINEER

Full Stack Developer & ML Engineer specializing in complex data extraction.

// My background in diagnosing critical systems at BMW taught me one thing: precision is non-negotiable. Today, I apply this rigor to software development and AI. Whether optimizing an API, improving UX, or fine-tuning a Computer Vision model, my goal remains the same: building tools that never fail.

> SELECTED_WORKS

01. InsureflowAi

1. XGBoost Filter 2. TATR Bounding Box 3. CamemBERT Vec HUMAN IN THE LOOP CONFIDENCE SCORE: 98.4% [ GREEN ] XLSX NORMALIZED

[01] PROBLEM STATEMENT

The insurance industry requires "Page-to-Data" extraction with surgical precision from highly heterogeneous PDF documents. Traditional OCR solutions prove incapable of maintaining logical structure across variable layouts (complex tables, nested hierarchies). This technological gap forces time-consuming, costly, and error-prone manual data entry, hindering the automation of critical business processes.

[02] TECHNICAL SOLUTION

Cascading model architecture: documents undergo initial filtering via XGBoost. Structural analysis (Layout Analysis) relies on a fine-tuned TATR model to accurately isolate tables. Hierarchy is reconstructed by a classification model factoring in layout and header semantics. Coverage lines are then normalized via a lightweight language model, specialized in French healthcare terminology. Everything is packaged in a standalone Windows executable: the entire pipeline runs locally on the client machine to ensure total privacy and eliminate cloud infrastructure costs.

[03] OPTIMIZATION / METRICS

The tool integrates a "Human-in-the-loop" mechanism with confidence scoring: each extracted cell is evaluated, and a color code guides the user to data requiring priority verification. Currently developed in Python, the solution is being migrated to a Rust / Tauri architecture. This rewrite aims to minimize the desktop application's memory footprint and ensure ultra-fast local inference performance, aligning with the sector's responsiveness requirements.

RAM PEAK : 4.5 GB → 2.2 GB (Optimized) → < 1.0 GB (Rust Target)

02. Layer0

LAYER0 // MAINTENANCE_DASHBOARD G-CODE_IN DB_POSTGRES MACHINE_WEAR Bambu P1S [ID: 04] WARNING: 85% LIFE DEPLETED FILAMENT_STOCK (PLA_MATTE) Daily Cons: 2.4kg LOG_MONITOR [OK] Parsed 240 G-Code instances. [OK] Calculated mass vs print time. [INFO] Synchronization standing by.

[01] PROBLEM STATEMENT

For a freelancer or small manufacturing workshop, managing a fleet of 3D printers quickly becomes a logistical challenge. True profitability is often blurry: between filament costs, invisible component wear (nozzles, belts), and power consumption, the cost price is frequently underestimated. Without a centralized tool, the craftsman loses precious time on administrative management instead of producing, directly impacting their growth.

[02] TECHNICAL SOLUTION

Containerized Nuxt 3 / Node.js architecture via Docker. The data flow relies on .gcode file ingestion: the user uploads their prepared file, triggering high-performance selective parsing. The algorithm analyzes machine instructions to extract critical metadata (prep time, print time, filament consumption) before injecting them into a PostgreSQL database. This process ensures precise indexing of each print job, transforming a simple manufacturing file into actionable accounting data.

[03] OPTIMIZATION / METRICS

The calculation dynamically integrates the complete economic model: machine depreciation, electricity, mechanical wear, and material cost to provide preventive alerts.

ASSET TRACKING : Depreciation based on target lifespan per 3D printer model (example → 5000h for Bambu Lab P1S)
INFRASTRUCTURE : VPS Ubuntu / Docker / PostgreSQL
> ACCESS_SYSTEM (PRE-PRODUCTION)

03. Trading SDK & Bridge

TradingView (Webhooks) BRIDGE // OS_WRAPPER (MAC/LINUX)
class MT5Bridge: def __enter__(self): self._connect_sockets() return self def __exit__(self, exc_type, exc_val, tb): # Guarantees socket release self._close_sockets() with MT5Bridge() as terminal: terminal.execute_trade()
MT5 TERMINAL SOCKETS [OK] FINANCIAL BROKERS

[01] PROBLEM STATEMENT

The native MetaTrader 5 API is restricted to the Windows ecosystem, lacks object abstraction, and exhibits chronic instability during prolonged executions. Developing cross-platform trading algorithms (macOS/Linux) requires a wrapper capable of securing data flows and normalizing interactions with the MetaTrader5 terminal.

[02] TECHNICAL SOLUTION

Development of an Algo-Trading SDK designed as a high-performance Wrapper Engine. The solution creates a logical "Bridge" allowing execution of complex strategies via structured Python objects. The architecture supports sending orders and retrieving real-time data while breaking free from OS constraints and the official MetaTrader5 library, transforming a low-level API into a modern and predictable development tool.

[03] OPTIMIZATION / METRICS

Priority was given to Resource Safety and network socket integrity. Implementation of rigorous Design Patterns, utilizing Context Managers (with) and magic methods (__enter__, __exit__) to ensure systematic opening and closing of connections.
This approach prevents any resource leaks or socket blocks, which are critical pain points during high-frequency or long-term trading sessions.

ARCHITECTURE: Python OOP / Webhooks / OS_Bridge
SECURITY: Context Hook __enter__ / __exit__
> VIEW_GITHUB_REPO

04. Equicares

Bras Avant-bras θ = 35° {"joint":"elbow", "θ":35} LLM_FEEDBACK_ENGINE > INGESTING VECTORS... > DETECTED_ISSUE: ELBOW_ANGLE Correction posturale : "Baissez la main droite."

[01] PROBLEM STATEMENT

Democratizing biomechanical analysis for riders via a centralized management platform. The challenge was to transform session photographs into actionable coaching advice, while centrally managing a stable's health, nutritional, and administrative tracking.

[02] TECHNICAL SOLUTION

SaaS platform integrating a hybrid AI stack. Posture analysis relies on YOLO (Pose Estimation) to extract articular Keypoints from imported snapshots. A vector algebra layer calculates the critical angulation θ, whose coordinates feed an LLM via a structured prompt to generate personalized coaching. In parallel, the backend manages nutritional calculation and care planning modules.

[03] OPTIMIZATION / METRICS

The solution validated interoperability between Computer Vision and GenAI on real-world use cases. While static analysis was rolled out, an R&D module on video analysis was conducted to explore dynamic movement before the project's shutdown. This post-mortem highlighted the critical importance of Product-Market Fit when dealing with a highly dense technical solution.

PIPELINE: YOLO Keypoints -> Angle Math -> LLM Prompt
KEY INSIGHT: Engineering != Product-Market Fit

> CAREER_SEQUENCING

bash - root@tpauwels:~/experience
[2024-PRESENT] ACTIVE_NODE: Independent Software Engineer
> Design of AI architectures specialized in data extraction (InsureTech sector).
> Lead Dev SaaS Equicares && OpenSource Layer0 / MT5 Bridge.
[2023-2024] INTEGRATION: Full Stack Developer & Data Analyst @ Hookto (Xefi group)
> Micro-services architecture and optimization of user experience (UX/UI) on business ERP.
> Management of asynchronous data flows via RabbitMQ and Redis.
> Stack: Nuxt 3, Ruby on Rails, RabbitMQ and Redis, PostgreSQL, Docker.
[2018-2021] DIAGNOSTIC: Expert After-Sales Technician @ BMW Bavaria Motor
> Analysis and resolution of breakdowns on critical embedded systems.
> Multiplex network diagnostics and high-precision mechanical troubleshooting.
> Technical right-hand: supervising adherence to manufacturer standards and recall campaigns.
> Managing complex cases in direct contact with BMW support engineers.
> Methodological rigor applied to high-responsibility environments.

> ACADEMIC_TIMELINE

bash - root@tpauwels:~/formations
[2023-BAC+4] STABLE_BUILD: Information Systems Expert - Data Major @ Xefi Academy
> Data Specialization: ML model development, Data Visualization, Statistics.
> Mobile and API Development (C#, Flutter, Swift, Python) && Linux Administration.
> Customer needs analysis && Software architecture design.
> Deployment automation and continuous integration (CI/CD).
[2022-BAC+4] STABLE_BUILD: Data Analyst @ OpenClassrooms
> Statistical modeling and implementation of Machine Learning algorithms.
> Data cleaning and preprocessing.
> Exploratory data analysis and visualization.
[2018] SUCCESSFUL_BUILD: Expert After-Sales Technician Qualification (TEAVA) @ BMW Group
> Valedictorian.
> Complex diagnostics and troubleshooting on critical embedded systems.
> Multiplex network analysis and technical incident management.
> Operational rigor applied to Premium vehicle maintenance.

> COMMUNICATION_MODULE

PINGING_ENDPOINTS...
EMAIL_ADDR: pauwelstanguy@protonmail.com
SECURE_LINE: 0686725419
_