Maxence is available for hire
Hire MaxenceMaxence Prevost
Verified Expert in Engineering
AI Engineer and Developer
Paris, France
Toptal member since April 10, 2026
Bio
Maxence is a senior AI/ML engineer who builds production-grade AI systems that go from research prototype to deployed product, fast. He has 9+ years of experience turning complex ML research into shipped products; real systems running in production. Maxence has delivered LLM-powered security agents, real-time audio transcription at 50 FPS on mobile devices, and end-to-end model quality management, from evaluation design to production monitoring.
Portfolio
Freelance
Artificial Intelligence (AI), Machine Learning, AI Agents, Agentic AI...
Codearena
Artificial Intelligence (AI), C, LangChain, Anthropic, LangGraph, OpenAI, DSPy...
Fretello
TensorFlow, Keras, Python, Convolutional Neural Networks (CNNs)...
Experience
- Python - 13 years
- Machine Learning - 9 years
- Deep Learning - 8 years
- Large Language Models (LLMs) - 4 years
- Prompt Engineering - 4 years
- Retrieval-augmented Generation (RAG) - 4 years
- OpenAI - 3 years
- AI Agents - 3 years
Preferred Environment
Python, TensorFlow, PyTorch, OpenAI, DSPy, LangChain, LangGraph, RAG Systems, Claude API, Software Engineering, SQL
The most amazing...
...thing I've built is a real-time polyphonic transcription system running at 50+ FPS on 2018 mobiles, with guitar technique analysis through AR feedback.
Work Experience
Senior AI/ML Engineer
2026 - PRESENT
Freelance
- Advised clients on LLM agent architecture, covering coordination patterns, reasoning strategies, and production reliability trade-offs.
- Designed evaluation pipelines for agentic AI systems, with per-slice metrics, LLM-as-judge checks, and regression benchmarks in CI.
- Built RAG prototypes over domain-specific corpora using hybrid semantic and keyword search with reranking to improve retrieval precision.
- Prototyped multi-agent workflows in LangGraph and DSPy.
- Implemented observability stacks with Langfuse for LLM applications, covering tracing, cost tracking, and quality monitoring in production.
Technologies: Artificial Intelligence (AI), Machine Learning, AI Agents, Agentic AI, Large Language Models (LLMs), Generative Artificial Intelligence (GenAI), Retrieval-augmented Generation (RAG), Prompt Engineering, Architecture, PostgreSQL, Software Development Lifecycle (SDLC), Automation, AI Automation, AI Agent Orchestration, Pandas, Hugging Face, RAG Systems, Workflow Automation, OpenAI API, AI Architecture, AI Chatbots, Vector Databases, Pydantic, RAG Architecture, REST APIs, Multi-agent Systems, Solution Architecture, Data Science, Machine Learning Operations (MLOps), LLM Reasoning, Knowledge Graphs, Claude Code, Python, Neural Networks, LangGraph, LangChain, OpenAI, Python 3, AI Design, AI Programming, Model Evaluation, Model Context Protocol (MCP), IT Strategy, Claude API, LLM Fine-tuning, Software Engineering, Langfuse, SQL, Orchestration, Agentic AI Systems, Agentic Frameworks, Pgvector, Prompt Optimization, Data Annotation, GEPA
Senior Machine Learning Engineer
2023 - 2025
Codearena
- Led end-to-end development of an LLM-powered security audit agent, built training datasets, and designed evaluation methodology.
- Designed agent workflows for threat modeling, vulnerability search, similarity search over past exploits, and static analysis integration, reducing manual audit effort.
- Developed a bug bounty recommendation engine matching 500+ security researchers to optimal codebases, improving audit coverage and researcher engagement.
- Built training datasets and designed evaluation methodology for LLM-based code analysis, with Langfuse-instrumented pipelines enabling iterative model quality improvement.
- Created a code comprehension tool adopted by security researchers to navigate unfamiliar codebases, reducing onboarding time for new audits.
Technologies: Artificial Intelligence (AI), C, LangChain, Anthropic, LangGraph, OpenAI, DSPy, Python, Programming, Machine Learning, AI Agents, Prompt Engineering, Large Language Models (LLMs), Generative Artificial Intelligence (GenAI), Agentic RAG Systems, Open-source LLMs, AI Chatbots, Agentic AI, AI Architecture, OpenAI API, Pinecone, FastAPI, Document Processing, Workflow Automation, RAG Systems, AI Model Training, Embedding Models, Fine-tuning, Natural Language Processing (NLP), Chatbots, Chatbot Conversation Design, Hugging Face Transformers, Synthetic Data Generation, Hugging Face, CI/CD Pipelines, Architecture, PostgreSQL, Software Development Lifecycle (SDLC), Automation, AI Automation, AI Agent Orchestration, Pandas, Technical Leadership, Vector Databases, Image Analysis, Pydantic, RAG Architecture, REST APIs, Multi-agent Systems, Solution Architecture, Data Science, Machine Learning Operations (MLOps), LLM Reasoning, Graph Databases, Knowledge Graphs, Pattern Recognition, Claude Code, Neural Networks, Python 3, AI Design, AI Programming, Model Evaluation, Model Context Protocol (MCP), IT Strategy, Claude API, LLM Fine-tuning, Software Engineering, Langfuse, SQL, Orchestration, Agentic AI Systems, Agentic Frameworks, Pgvector, Prompt Optimization, A/B Testing, Data Annotation, GEPA
Deep Learning Engineer
2020 - 2023
Fretello
- Led the R&D of a real-time polyphonic guitar transcription system using a multi-head CNN architecture, serving thousands of active users in production on iOS and Android.
- Collected and curated 120 hours of audio training datasets. Designed annotation workflows, data quality checks, and augmentation pipelines to improve model robustness.
- Optimized inference and signal processing pipelines to achieve around 50 frames per second (FPS) on mobile hardware, enabling real-time feedback for guitarists during live practice sessions.
- Evaluated model performance through iterative experimentation cycles, improving transcription accuracy while maintaining real-time low-latency requirements on mobile.
Technologies: TensorFlow, Keras, Python, Convolutional Neural Networks (CNNs), Audio Processing, Signal Processing, Programming, Artificial Intelligence (AI), OpenCV, Machine Learning, Deep Learning, Generative Artificial Intelligence (GenAI), AI Architecture, Computer Vision, FastAPI, Amazon Web Services (AWS), Document Processing, AI Model Training, Embedding Models, Fine-tuning, Synthetic Data Generation, CI/CD Pipelines, Architecture, Audio, Software Development Lifecycle (SDLC), Automation, AI Automation, Pandas, Hugging Face, Open-source LLMs, Technical Leadership, Image Segmentation, REST APIs, Solution Architecture, Data Science, Machine Learning Operations (MLOps), Pattern Recognition, Image Processing, Neural Networks, PyTorch, Python 3, AI Design, Model Evaluation, IT Strategy, Software Engineering, Grafana, SQL, A/B Testing, Data Annotation
Deep Learning Research Engineer
2018 - 2020
CloudLinux
- Built a PHP vulnerability scanner using deep learning (CNNs and RNNs) to detect security flaws in source code, improving detection accuracy over rule-based approaches by 35%.
- Designed and trained sequence-to-sequence models on large PHP codebases to identify SQL injection, XSS, and command injection patterns at scale.
- Developed a data pipeline to parse, tokenize, and preprocess PHP source code into model-ready representations, handling over 500,000 code samples.
- Collaborated with the cybersecurity team to validate model outputs and integrate the scanner into CloudLinux's existing security toolchain.
Technologies: Python, TensorFlow, Recurrent Neural Networks (RNNs), Long Short-term Memory (LSTM), Convolutional Neural Networks (CNNs), Programming, C++, Artificial Intelligence (AI), Keras, Machine Learning, Deep Learning, Amazon Web Services (AWS), Document Processing, AI Model Training, Embedding Models, Fine-tuning, Natural Language Processing (NLP), Synthetic Data Generation, Architecture, PostgreSQL, Software Development Lifecycle (SDLC), Automation, AI Automation, Pandas, Workflow Automation, Technical Leadership, REST APIs, Solution Architecture, Data Science, Graph Databases, Knowledge Graphs, Pattern Recognition, Neural Networks, PyTorch, Python 3, AI Design, Model Evaluation, Software Engineering, Grafana, SQL, A/B Testing, Data Annotation
Software Engineer, Computer Vision
2017 - 2018
TRAI Systems
- Developed real-time computer vision modules in C and C++ for embedded systems, enabling object detection and tracking on resource-constrained hardware.
- Implemented and optimized image processing pipelines using OpenCV, reducing processing latency for live video streams by 40%.
- Designed and integrated custom vision algorithms for industrial quality control, reducing manual inspection time by 60% on the production line.
- Contributed to the full software development lifecycle from requirements analysis through testing and deployment on embedded Linux platforms.
Technologies: C++, OpenCV, Computer Vision, Python, Programming, Machine Learning, Deep Learning, Architecture, Software Development Lifecycle (SDLC), Point Clouds, 3D, Image Analysis, Image Segmentation, Solution Architecture, Data Science, Pattern Recognition, Image Processing, Neural Networks, Python 3, AI Design, Software Engineering, SQL, A/B Testing, Data Annotation
Experience
Real-time Polyphonic Guitar Transcription System
Designed and built a real-time polyphonic guitar transcription system from the ground up, deployed in production on iOS and Android for the music education app Fretello, with thousands of active daily users.
The core system uses a multi-head CNN architecture to detect simultaneous notes and chords from raw audio captured via the device microphone, running inference at 50+ FPS on 2018-era mobile hardware. I led the full ML lifecycle: dataset collection and curation (120+ hours of labeled guitar audio), annotation workflow design, augmentation pipelines, model training and evaluation, and mobile optimization via quantization and ONNX export.
Beyond transcription, I engineered a signal-processing pipeline to extract guitar techniques (bends, vibrato, hammer-ons) and integrated augmented-reality visual feedback to help learners improve their playing in real time. The pipeline achieved latency under 20 milliseconds end-to-end on-device, making interactive feedback viable for live practice.
Stack: Python, TensorFlow, Keras, C++, Core ML, ONNX, librosa, and custom DSP
The core system uses a multi-head CNN architecture to detect simultaneous notes and chords from raw audio captured via the device microphone, running inference at 50+ FPS on 2018-era mobile hardware. I led the full ML lifecycle: dataset collection and curation (120+ hours of labeled guitar audio), annotation workflow design, augmentation pipelines, model training and evaluation, and mobile optimization via quantization and ONNX export.
Beyond transcription, I engineered a signal-processing pipeline to extract guitar techniques (bends, vibrato, hammer-ons) and integrated augmented-reality visual feedback to help learners improve their playing in real time. The pipeline achieved latency under 20 milliseconds end-to-end on-device, making interactive feedback viable for live practice.
Stack: Python, TensorFlow, Keras, C++, Core ML, ONNX, librosa, and custom DSP
Bug Bounty Recommendation Engine for Security Researchers
I designed and shipped a recommendation engine that matches security researchers to bug bounty programs on the Immunefi platform. The goal was to increase audit coverage on newly launched programs and drive researcher engagement by surfacing the codebases each researcher is most likely to find bugs in.
Feature development drew from three sources: researcher history (past submissions, severities, specialties), program metadata (stack, protocol type, complexity, payout tiers), and behavioral signals from the platform. The ranker evaluated over 500 active researchers against the full live program catalog, with scores feeding directly into the researcher-facing product.
Performance was evaluated using a held-out slice of real matches and a retrospective check on known high-signal researcher-to-program pairings, since a simple “did they submit” label is too sparse to be useful on its own.
End-to-end ownership covered data ingestion from the platform database, the feature store, model training, and the serving API consumed by the product team. An offline evaluation harness ensured every model change passed before going live, catching regressions before they impacted real researchers.
Feature development drew from three sources: researcher history (past submissions, severities, specialties), program metadata (stack, protocol type, complexity, payout tiers), and behavioral signals from the platform. The ranker evaluated over 500 active researchers against the full live program catalog, with scores feeding directly into the researcher-facing product.
Performance was evaluated using a held-out slice of real matches and a retrospective check on known high-signal researcher-to-program pairings, since a simple “did they submit” label is too sparse to be useful on its own.
End-to-end ownership covered data ingestion from the platform database, the feature store, model training, and the serving API consumed by the product team. An offline evaluation harness ensured every model change passed before going live, catching regressions before they impacted real researchers.
LLM-powered Security Audit Agent
I built an LLM-based agent that autonomously reviews smart contract code for security vulnerabilities.
As a senior AI/ML engineer, I designed and owned the full pipeline end-to-end: training data collection and curation from public bug bounty reports; prompt engineering and fine-tuning strategies for vulnerability detection; RAG-based context injection for contract-specific knowledge; and a structured evaluation framework to measure autonomous coverage.
The agent achieves around 30% autonomous vulnerability coverage on real-world smart contracts, surfacing critical issues such as reentrancy, integer overflow, and access control flaws without human intervention.
Key challenges included handling large and complex codebases, minimizing false-positive rates while maintaining recall, and designing an evaluation pipeline robust enough to track model improvements across iterations.
Stack: Python, LangChain, LangGraph, OpenAI/Anthropic APIs, RAG, DSPy, and custom evaluation harness
As a senior AI/ML engineer, I designed and owned the full pipeline end-to-end: training data collection and curation from public bug bounty reports; prompt engineering and fine-tuning strategies for vulnerability detection; RAG-based context injection for contract-specific knowledge; and a structured evaluation framework to measure autonomous coverage.
The agent achieves around 30% autonomous vulnerability coverage on real-world smart contracts, surfacing critical issues such as reentrancy, integer overflow, and access control flaws without human intervention.
Key challenges included handling large and complex codebases, minimizing false-positive rates while maintaining recall, and designing an evaluation pipeline robust enough to track model improvements across iterations.
Stack: Python, LangChain, LangGraph, OpenAI/Anthropic APIs, RAG, DSPy, and custom evaluation harness
Real-time AR Guitar Learning
https://fretello.com/news/mirror-revolutionizing-guitar-learning-with-augmented-reality/At Fretello, I worked on Mirror, a feature that turns a phone’s front-facing camera into a live guitar tutor. The system recognizes the guitar and its fretboard orientation from the camera feed, then draws finger positions and strumming cues straight onto the instrument while you play.
The hard part was the computer vision. Detecting and tracking the fretboard frame by frame, fast enough to feel real-time on mobile, across bad lighting and odd angles, without the lag that would break the overlay. I built the detection and tracking side and the pipeline that keeps the cues aligned as the guitar moves.
The hard part was the computer vision. Detecting and tracking the fretboard frame by frame, fast enough to feel real-time on mobile, across bad lighting and odd angles, without the lag that would break the overlay. I built the detection and tracking side and the pipeline that keeps the cues aligned as the guitar moves.
Computer Vision Engineer
MCQ-Scan is a French computer vision and 3D scanning specialist based in Lille that develops autonomous mobile robots for retail and logistics. Their primary retail solution is StoreBot, a fully autonomous robot designed to analyze store aisles and detect anomalies such as stockouts, misplaced products, and pricing errors or missing labels.
Education
2010 - 2016
Master's Degree in Computer Science
Institut Supérieur d'Electronique et du Numérique - Lille, France
Skills
Libraries/APIs
TensorFlow, PyTorch, Keras, OpenCV, OpenAI API, Hugging Face Transformers, REST APIs, Claude API, Pydantic, Pandas
Tools
Claude Code, Grafana
Languages
Python, Python 3, SQL, C++, C
Frameworks
DSPy, LangGraph, Agentic Frameworks
Paradigms
Synthetic Data Generation, Automation, Model Context Protocol (MCP)
Platforms
Langfuse, LangSmith, Amazon Web Services (AWS)
Storage
PostgreSQL, Graph Databases
Other
OpenAI, Anthropic, LangChain, Neural Networks, Programming, Artificial Intelligence (AI), Convolutional Neural Networks (CNNs), Computer Vision, Retrieval-augmented Generation (RAG), Large Language Models (LLMs), Prompt Engineering, AI Agents, Machine Learning, Deep Learning, Generative Artificial Intelligence (GenAI), Agentic RAG Systems, Open-source LLMs, AI Chatbots, Agentic AI, AI Architecture, Document Processing, Workflow Automation, RAG Systems, AI Model Training, Embedding Models, Fine-tuning, Natural Language Processing (NLP), Chatbots, Chatbot Conversation Design, Hugging Face, Architecture, Vector Databases, Image Analysis, Image Segmentation, RAG Architecture, Multi-agent Systems, Solution Architecture, Data Science, LLM Reasoning, Pattern Recognition, Image Processing, AI Design, AI Programming, Model Evaluation, Software Engineering, Orchestration, Agentic AI Systems, Pgvector, Prompt Optimization, GEPA, Pinecone, FastAPI, Technical Leadership, Software Development Lifecycle (SDLC), AI Automation, AI Agent Orchestration, Audio, Machine Learning Operations (MLOps), Knowledge Graphs, IT Strategy, LLM Fine-tuning, A/B Testing, Data Annotation, Audio Processing, Signal Processing, Recurrent Neural Networks (RNNs), Long Short-term Memory (LSTM), Supabase, CI/CD Pipelines, Point Clouds, 3D, Virtual Reality (VR)
Collaboration That Works
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
1
Share your needs
Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2
Choose your talent
Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3
Start your risk-free talent trial
Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.
Top talent is in high demand.
Start hiring