Antoine A.

landing.arkhivesstudio.com • Co-Founder & CTO at Arkhives Studio, spearheading the development of Forge, an AI agent-orchestrated production platform for games, film, and TV. • Focused on enhancing production velocity while ensuring enterprise-grade trust through consent-first data and human-in-the-loop approvals. • Developing Arkhives Originals to validate the pipeline, achieving rapid iteration targets and bridging the gap from concept to production-ready. • Trained tool-using agents with OpenPipe ART reinforcement framework for MCP tool/skill execution; improved tool-use quality on internal evals from ~60% → ~88% (≈ +28 pp, ~47% relative reduction in tool hallucinations). • Designed an evaluation-gated promotion policy: models must exceed ≥80% success on targeted tool/skill suites before being promoted into broader evaluation and release candidacy. • Built a judge–teacher–student distillation pipeline: teacher generates synthetic datasets from real traces; judge scores student behavior; student is trained + re-evaluated against a golden dataset to prevent drift. • Produced synthetic evaluation datasets in ~1,000 and ~10,000 example/trajectory tiers to scale eval coverage and harden tool-use behavior prior to promotion. •Quantized FP16 models down to FP4/NVFP4 (NVIDIA/TE-class) and INT4 using NVIDIA CUDA tooling; maintained performance thresholds via gated evals. • Reduced VRAM footprint from ~30 GB → <10 GB (≈ ≥67% reduction) through quantization + deployment tuning, enabling broader deployment options and higher concurrency. • Authored FastMCP servers and clients to connect internal tools safely and consistently; standardized tool schemas and orchestration patterns for agent reliability. • Implemented model snapshot promotion and checkpoint durability across S3, private Hugging Face Hub checkpoints, and an additional private cloud provider, ensuring recoverability and reproducibility.

• Led the GenAI program portfolio, focusing on ChatGPT enablement and governance. • Conducted AI architecture reviews in collaboration with CyberSecurity, DevOps, and BI teams. • Developed GenAI security guardrails and initiated the MCP server project for internal tool integration. • Rolled out AI-assisted code review and piloted AI code generation to enhance development workflows. • Trained tool-using agents with OpenPipe ART reinforcement framework on MCP skills; implemented reward shaping + failure taxonomy to improve tool-call reliability and reduce invalid actions. • Built eval-gated promotion pipeline: models promoted when they exceed ≥80% skill/tool success threshold across offline suites; automatically routed to merge/swap decisions for production flows. • Designed judge–teacher–student distillation workflow: teacher generates synthetic data from real traces; judge scores student outputs; synthetic dataset used for both training and validation with a real “golden” dataset as ground truth anchor. • Produced task-specific datasets (you mentioned a “vow’s dataset”—if that’s proprietary, rename it generically like “tool-preference dataset” or “behavioral preference dataset”). • Quantized FP16 models down to FP4 (NV/Transformer Engine) and INT4 using NVIDIA CUDA tooling; validated quality retention against tool-skill evaluation suites and enforced regression thresholds pre-promotion.

- Trained tool-using agents with OpenPipe ART reinforcement framework for MCP tool/skill execution; improved tool-use quality on internal evals from ~60% → ~88% (≈ +28 pp, ~47% relative reduction in tool hallucinations). - Designed an evaluation-gated promotion policy: models must exceed ≥80% success on targeted tool/skill suites before being promoted into broader evaluation and release candidacy. - Built a judge–teacher–student distillation pipeline: teacher generates synthetic datasets from real traces; judge scores student behavior; student is trained + re-evaluated against a golden dataset to prevent drift. - Produced synthetic evaluation datasets in ~1,000 and ~10,000 example/trajectory tiers to scale eval coverage and harden tool-use behavior prior to promotion. - Quantized FP16 models down to FP4 (NVIDIA/TE-class) and INT4 using NVIDIA CUDA tooling; maintained performance thresholds via gated evals. - Reduced VRAM footprint from ~30 GB → <10 GB (≈ ≥67% reduction) through quantization + deployment tuning, enabling broader deployment options and higher concurrency. - Implemented production agents using LangChain and LlamaIndex, integrating structured tool execution through MCP. - Authored FastMCP servers and clients to connect internal tools safely and consistently; standardized tool schemas and orchestration patterns for agent reliability. - Built agent-to-agent communication via A2A protocol, including a library (CyberRelay) to support multi-agent workflows and interoperability. - Ran fine-tuning/quantization workflows on a 4-node local Spark cluster connected via InfiniBand (ConnectX-7); optimized for repeatable promotion runs. - Implemented model snapshot promotion and checkpoint durability across S3, private Hugging Face Hub checkpoints, and an additional private cloud provider, ensuring recoverability and reproducibility.

- Content Management System Development: Collaborated in the co-development of a robust Content Management and Authoring System, enhancing stability and facilitating the efficient sharing of WDI design content. - Chatbot Development: Engineered a functional chatbot with AI-driven recommendation capabilities for a beta release, significantly improving issue resolution for cast members. (Gen AI/LLM) - AI Character Interaction System: Spearheaded the creation of a beta version of an AI character interaction system, designed to enhance cast member and guest engagements. (Spatial Computing/XR) - User Interface Design: Crafted intuitive and user-friendly interfaces for both the chatbot and content management systems, focusing on ease of use and accessibility. - System Integration: Seamlessly integrated innovative solutions with existing organizational frameworks, ensuring compatibility and functionality. - Quality Assurance: Led rigorous testing protocols to ensure the performance and reliability of new systems, maintaining high standards of quality. - Stakeholder Engagement: Organized and conducted stakeholder workshops to gather actionable feedback and foster collaborative development. - Technical Documentation: Authored comprehensive guides and documentation to support system implementation and usage. - Project Handover: Successfully transitioned projects to WDI R&D and technology stakeholders, ensuring continuity and operational efficiency. - Post-Implementation Support: Provided ongoing support and training post-implementation, enhancing user adoption and system effectiveness.

• Proficient in AI technologies, notably ChatGPT APIs, and adept at leveraging them for app development, including a notable PDF summarizer app using Node.js and Python/Flask. • Displayed outstanding problem-solving skills and attention to detail in handling complex tasks such as the implementation of a custom navigation feature for a Healthcare client’s Prototype using React Native/TypeScript. • Praised for exceptional coding practices, including clean and understandable code, and refactoring of context and providers from class-based to functional ones on a Retail React Native/TypeScript app project. • Demonstrated strong open communication and feedback skills, which are important for collaboration in high-impact projects. • Commended for the ability to simplify complex problems, evident through successful mentorship and guidance in team huddles. • Recognized for quick adaptability and contribution to new teams and projects, shown through immediate value addition and proactive bug-fixing within tight timelines. • Recognized for world-class communication, a vital aspect of transparency, accountability, and ethics in project development. • Orchestrated the development of a high-impact Adobe AEM integration showcase Prototype app, leveraging technologies like React Native, TypeScript, React-Query, Context API, and GraphQL. This ambitious project was unveiled at the prestigious Adobe Summit Conference in Las Vegas, NV, showcasing my ability to deliver impressive results under a strict two-week deadline.

• Implemented higher poly models in a React Native app using the new React Native Architecture. • Utilized C++ to optimize app performance and achieve seamless integration with existing codebase. • Collaborated with cross-functional teams, including designers and product managers, to ensure timely delivery of features and updates. • Mentored junior engineers in React Native, C++, and XR best practices.

• Led cross-functional teams of 8 (engineers, SCRUM masters, product designers) to deliver enterprise and startup app contracts. • Guided technical decision-making in collaboration with clients, design, and product teams. • Designed training methodologies to enhance junior engineers’ understanding of modern technology. Training Initiatives for OF Developers: • VR/AR development in React Native with ViroMedia • Cloud applications and API development with Golang • Native mobile module development with Golang • Data engineering with Python/Golang Project Highlights: • Developed a machine learning algorithm and data pipeline using Node.js and Algolia for a FinTech app’s recommender API. Client Engagements: ShareMy.Health: • Advised on React Native best practices • Collaborated with cloud architects and business stakeholders to set up CI/CD pipelines in AppCenter for app deployments to stores • Translated React features into React Native in collaboration with the Web/React team Anthem: • Advised on React Native best practices • Contributed fixes and updates to ensure the React Native/TypeScript codebase was 508 (Accessibility) compliant

Lead a team of 15 (Software Engineers, QA, DevOps, Data Engineers) to maintain monolith MERN-built SAAS product Managed the development of new features for Industry-Leading clients like: Facebook, SC Johnson, BAYER, and more. Guided Software Architectural decision-making based on meetings with CEO, CTO, CSO, VP of Product Strategy, Directors of: Data Science, Product Ops, and Client Success. Lead group with Directors of Data Science & Product Ops on creating training methodologies to enhance our teams' understanding of modern technologies and methodologies like Natural Language Processing, Data Engineering for endpoints, Micro- Architectures for Frontend & Backend, AWS Deployment, Git Strategy, and Industry Best Practices for Component-Driven-Development, Behavior-Drive-Development and Test-Driven-Development. Acted swiftly to maintain the Development team's quarterly goals and velocity after an abrupt CTO departure. Architected the future Tech Stack for SaaS product to move from a fragile, rigid software design to a modular, elastic software design.

Lead engineer for a Childcare management application

Improved App Store rating from 2.3 to 3.5 Lead redevelopment of PartyCity mobile app on a team of 2 engineers under significant time crunch.