Tomy Spagnoletti Duval

Working across Dubai and La Rochelle for Rhinos Groupe and Rhinos ME, with a mandate centered on company-wide Codex adoption, technical enablement, and R&D execution.

Company

• Rhinos is an independent French group active since 1998, focused on B2B relationship marketing, client loyalty, sales-network activation, referral mechanics, operational support, logistics, and dedicated digital execution.
• While Rhinos ME anchors the Dubai presence, the operational center of gravity remains Rhinos Groupe in La Rochelle, where the group concentrates its main execution force with 100+ employees on site inside a larger organization spanning both hubs.
• Rhinos is a market leader in its category, serving major accounts including Saint-Gobain, Covéa, Volkswagen Group, Loxam, and Groupe Carrefour, among others.

Codex rollout

• Led the rollout of Codex and ChatGPT across the company, alongside other AI tools selected for specific team workflows, turning AI adoption into an operational capability rather than a side experiment.
• Designed and delivered role-specific training, workshops, and hands-on enablement sessions for multiple teams, adapting the rollout to the needs of each function instead of forcing one generic AI workflow on everyone.
• Served as a trusted technical partner for teams adopting AI in their daily SDLC, from workflow design to hands-on enablement, operational readiness, and safe rollout.

Systems & execution

• Built internal tools with Codex for different parts of the company, focusing on practical leverage across operations rather than isolated demos.
• Currently building an in-house OpenClaw-like tool as a company-wide agent layer for Rhinos. It is designed to support external customer operations while also assisting internal teams across development, IT, security, DevOps, accounting, and broader business workflows.
• The target architecture is an AI system connected to company knowledge across GitHub repositories, Confluence documentation, and internal technical context, so the organization can rely on one global assistance layer instead of fragmented AI usage.

In short, I operate as the AI technical partner helping Rhinos turn frontier models into real company-wide execution.

Building Vivicy, an AI coding workflow platform that transforms a GitHub repository into a reactive graph of mini-spec blocks that orchestrate Codex as the coding agent behind structured autonomous work.

Product

• Designed the board-centric control plane behind that graph, connecting mini-spec blocks, Codex runs, execution loops, GitHub workflows, and delivery decisions.
• Built the user-facing product around graph-based mini-spec boards, GitHub repo connection, typed block dependencies, chat-assisted board edits, and Codex run controls. It includes configurable pipeline stages, live logs, diff and artifact review, drift and external-resource warnings, snapshots, import/export, and secure secrets. It keeps a locked LLM runtime contract on @openai/codex-sdk.
• Built the product with a clear dogfooding target: use Vivicy to build and evolve Vivicy, so the platform is tested against the same agent workflows it is meant to support for other teams.
• The goal is to make AI-assisted development maintainable: track intent, affected files, dependencies, run history, failures, reviews, and drift instead of treating generated code as a black box.

Harness engineering

• Built the repo around strong harness engineering: agents get the context, constraints, checks, and diagnostics they need before and after a run. They do not operate from a loose prompt while everyone hopes the output is safe.
• Treated repository knowledge as something the system can verify: docs, generated inventories, route classifications, runtime contracts, and failure prompts keep the codebase understandable for the next agent run.
• Added anti-drift controls around protected files, route and environment inventories, RLS checks, repo hygiene, and corrective failure outputs, so changes break loudly when the system no longer matches its contracts.

Proof system

• Built strict E2E gates and runtime contracts so agent-generated changes are backed by browser evidence, performance budgets, and fail-fast diagnostics.
• On critical surfaces, E2E is explicitly designed to reduce manual regression QA. The repo enforces `governed_e2e_coverage = 100`, real performance budgets, worker-safe execution, and fail-fast runtime policies so autonomy is backed by evidence, not optimism.
• The product is built for the real failure modes of AI coding: incomplete specs, changed dependencies, broken interfaces, merge conflicts, stale assumptions, and runs that need a clear audit trail.

In short, Vivicy is where I build the operating layer that gives Codex safe autonomy on a real codebase through graph-based context and run history, while keeping intent, traceability, and engineering control intact.

Built WPLLMT in 10 days for my own operational needs: an open-source, privacy-first WordPress migration tool combining one-way SFTP sync, full-refresh MySQL migrations, local-only browser state, and an integrity checker for server-to-server moves.

Shipped as a product split across the app and its public website. GitHub / Live site.

Built Ton Soutien Psy, a search-first platform designed to help patients find "Mon Soutien Psy" accredited psychologists faster than the official directory. It brings stronger UX, deeper page coverage, and a more serious SEO foundation.

Product & search

• Designed the full search and discovery layer across specialties, audience, teleconsultation, city filters, alphabetical directory pages, city landing pages, and practitioner profiles.
• Built the product on Next.js App Router + Tailwind CSS 4, with route patterns and interface flows tuned for a fast, low-friction directory experience.

SEO

• Structured the site as a programmatic SEO and Generative Engine Optimization (GEO) system with dedicated city, specialty, directory, and profile surfaces. It also includes canonical metadata, dynamic sitemap coverage, llms.txt guidance for assistants and crawlers, and AI-enriched practitioner and city content.
• The project is already driving 2,000+ visits per day while the SEO layer is still only around 5% deployed.

Data pipeline

• Built the ingestion pipeline from the Mon Soutien Psy / Ameli ecosystem, using proxies, a local SQLite staging layer, guarded Postgres loads, diff reports, and explicit confirmation before writes.
• Added enrichment flows for specialties, languages, and profile descriptions, with AI-assisted generation and verification passes to keep outputs constrained and consistent at scale.

Infra & quality

• Modeled the dataset in Prisma + Postgres, added search and ranking primitives with pg_trgm, geospatial proximity logic, and precomputed city-level aggregates.
• Backed the product with unit, UI, integration, and E2E coverage to keep a content-heavy search product stable as the index keeps expanding.

In short, I built Ton Soutien Psy to give more visibility to psychologists who make the effort to be part of the Mon Soutien Psy program. It gives patients in France the fast, clear, trustworthy search experience that was missing: careful UX, clean UI, and useful, up-to-date information.

Developed P99, a Solana-first automated trading system designed as a distributed set of agents handling ingestion, signal selection, execution, and position management.

Architecture

• Designed a multi-agent runtime with dedicated services for stream ingestion, token selection, trade execution, and watcher logic.
• Structured inter-agent communication around Redis Streams, with mint-level locking, in-flight deduplication, and lifecycle phases for observability and idempotence.
• Orchestrated the whole runtime with PM2 and prepared durable persistence with Prisma + PostgreSQL.

Trading stack

• Built the first execution layer on Solana, with upstream event ingestion around Pump migrations and order execution through Jupiter Ultra.
• Added operator surfaces with a Next.js + Tailwind CSS 4 dashboard foundation and an Ink terminal UI for live runtime controls.

Engineering

• Built the repository as a Bun + strict TypeScript monorepo with shared packages, agent workspaces, tests, and ops tooling.
• Treated security and operational resilience as first-class concerns due to wallet key handling and real on-chain execution.

Built the BrainRoad Labs website from scratch with Astro JS. I designed a custom scroll engine with inertia analysis to ensure smooth, intuitive navigation — especially on Apple devices — and delivered a fully responsive, high-performance UI/UX.

The site you are reading now is the current V2, rebuilt from that first version with a sharper editorial direction and a more focused portfolio structure.

V1 screenshots

Built BETSY as a prototype for an AI-powered virtual language teacher that speaks in real time using GPT-4o and LiveKit, designed to explore what a more natural, instructor-like language learning experience could feel like.

Engineered CheckAI, a proof-of-concept change detection tool using the “The Change You Want to See” co-attention model (WACV 2023) to automate before/after image comparison for short-term rentals. I integrated the full inference pipeline via Flask and served a responsive UI with bounding-box overlays.

To meet hardware constraints, I adapted the training and inference flow to run entirely on CPU — including Apple Silicon — by bypassing CUDA dependencies, optimizing model loading, and tweaking detection thresholds and visualizations.

The result: on-demand visual diffing with pixel-accurate masks and seamless UX on low-cost infrastructure.

Rebuilt IK-Generator (v2) App, ten years after developing its first version. Migrated a monolithic Symfony app with MySQL, Elasticsearch, and OSRM — previously installed bare-metal — into a fully containerized architecture using Docker and Docker Compose.

Implemented secure server automation (Debian hardening, SSH lockdown, firewall, fail2ban...), streamlined data pipelines for database imports and ES population, and created scripted deployment flows for both dev and prod environments on Scaleway.

Built the end-to-end GenAI experience for Jean-Paul Gaultier's ecommerce site. Users pick a Parisian phrase (available in 7 languages), record a short mouth clip, and the system isolates, beautifies, and composites the lips onto AI-generated skies, publishing brand-safe videos to a live “wall of mouths”.

At each key step, Xano orchestrated the flow and sent transactional emails via SendGrid to keep users informed.

Role & scope

• Lead engineer: front-end JavaScript, back-end, GenAI/ML pipeline (ComfyUI), DevOps, and moderation tooling.
• Team: UX/UI designer + one front-end collaborator; I owned the pipeline, infra, and all of the app logic.

Stack

• GenAI / ML: ComfyUI with custom SDXL model, custom LoRA(s) for lipstick styling (trained on synthetic dataset, released on CivitAI), AnimateDiff for frame-to-frame consistency, SAM + Grounding DINO for segmentation/detection, ControlNet, CLIP (pos/neg), KSampler, FILM VFI for interpolation, VAE, video combine...
• Infra: AWS EC2 (ComfyUI in Docker), S3 for input/output storage.
• Backend: Xano (job orchestration, APIs, DB in Postgres, callback handling, SendGrid integration).
• Front: vanilla JavaScript experience embedded in Jean-Paul Gaultier's site; mobile & desktop camera capture.
• Content: backgrounds via Midjourney, Jean-Paul Gaultier's voiceover via ElevenLabs.
• Admin: custom Vue.js back-office for moderation and wall management.

See the detailed ComfyUI workflow Optional 10-step technical pipeline for reviewers who want the details

1. Fetch video from S3 → drop 1 frame out of 2 to reduce RAM, since one user video produced 3 outputs (HD+FHD landscape for the site, and FHD portrait for social).
2. Generate new video via KSampler with prompts (pos/neg), ControlNets, the source video, and AnimateDiff to enforce temporal consistency.
3. Segment the new video.
4. In parallel, generate a second “lipstick-only” red-lips video, used as a guiding mask to help SAM + Grounding DINO perform more accurate segmentation.
5. Apply this mask to the main custom video.
6. Upscale useful frames from 1024x576 to HD and FHD variants.
7. Composite lips onto AI skies.
8. FILM VFI interpolation → 60 fps for smoothness.
9. Mix original audio back into the final render.
10. Push 3 new videos (HD, FHD, social) to S3 + Postgres/Xano, triggering callbacks and user notifications via SendGrid.

Engineering highlights

• Designed a multi-branch ComfyUI pipeline combining AnimateDiff, ControlNets, segmentation feedback loops (red-lips masking) and more.
• Balanced RAM + GPU limits by frame-sampling + smart upscale, while still hitting high-res deliverables (example below).
• Automated 7-language UX with transactional emails for transparency (recorded → processing → published).
• Built a moderation back-office (Vue.js) with review queues, “On the Wall” toggles, and safe-publish controls.
• Containerized inference on EC2 for reproducibility and scale.

Result

A visually striking, multilingual UGC experience that transforms noisy webcam clips into branded AI videos, manages moderation, and delights users with live publication on the “wall of mouths”.

Takeaway

In short, DoYouSpeakGaultier? was an unprecedented generative-AI marketing operation in France, built when the ecosystem was still emerging and the tooling was highly experimental. For me, it was a rare full-stack challenge: front-end capture, backend orchestration, AWS inference infrastructure, moderation tooling, and a 17-model chained GenAI pipeline. It was technically broad, difficult, and deeply formative.

Screenshots

Live Generation Example

• Pivot from Ko Social Network (below) to focus on AI-powered content monetization.
• Developed Shaaare, a social network where photo & video posts were automatically evaluated by GPT-4o, rewarding creators in XKO tokens (25–100 XKO per post) directly on blockchain.
• Implemented anti-spam measures (3–5 posts/day limit, verified users) and automatic rejection of low-quality content (score < 25).
• Built and released three full versions, published on the Google Play Store, blocked by the Apple App Store due to in-app crypto monetization rules.
• Achieved JEI (Jeune Entreprise Innovante) status in France.
• Raised nearly €1M from 100+ investors across both Ko & Shaaare projects.

Resources:

• – Presentation Video
• – Archived Website & Whitepaper

Latest Store Screenshots

Built and led a Web3 social network with a team of front-end and back-end developers, UI/UX designers, and a marketing squad. The project aimed to give content ownership back to users through NFT-based posts and a native cryptocurrency (XKO).

• Decentralized Ownership: Users could mint photo and video NFTs directly from their devices and manage them via a non-custodial in-app wallet.
• Seamless UX + Security: Implemented ERC-20 meta-transactions on Polygon and Ethereum, enabling gasless transactions with transparent signature handling.
• Native Token Economy: Designed Like+Ko (tip-based likes with customizable XKO amounts) and Pay-to-View content monetization.
• Crypto Features: Enabled peer-to-peer XKO transfers and NFT minting with frictionless mobile experience.
• Recognition: Invited to CES 2023 with a "Spotlight Tech for Good" mention, leading to high-level discussions with Jamie Holcombe (USPTO CIO).
• Social App Experience: Positioned as a blend of TikTok + Instagram, with support for text, audio, photo, and video posts — NFTs reserved for photo/video content.

Despite full feature completion and many shipped versions, Apple Store in-app monetization restrictions and fundraising challenges in France led to the project pivoting into Shaaare (above).

Latest Store Screenshots

• TSI (Tout Sauf l’Isolation): Developed a nationwide ERP in FileMaker to manage insulation work projects, fully compliant with ADEME’s regulatory requirements.
• LiveCoworking: Built a web platform enabling restaurants to convert unused space into coworking areas during off-service hours, designed with a European expansion in mind. Activity halted due to COVID.
• LiveCo.: Created a real-time Mapbox + Jitsi application where users could generate and join video-conference rooms directly on a shared map — fostering spontaneous collaboration and discussion during the pandemic.

• Served as principal instructor and training manager for a three-year Web & Mobile Development program at Webstart Paris, a hands-on digital school in central Paris focused on practical and innovative pedagogy.
• Oversaw three cohorts (Dev 1, Dev 2, Dev 3). I managed the full curriculum and coordinated the teaching team, including specialists in C/C#, back-end development, and UX/UI design.
• Personally delivered general technology foundation and front-end development courses (HTML, CSS, JavaScript) to aspiring developers over two and a half years.
• Fostered a collaborative learning environment, driving both technical mastery and soft skills within a practice-oriented educational culture.
• This remains the foundation of how I teach technical adoption today: clear structure, live practice, and hands-on support rather than abstract theory.

Developed and scaled a SaaS platform for calculating and managing French mileage allowances (Indemnités Kilométriques), later handed over to my wife (in 2024).

• Unique core feature: generated mileage allowance tables directly from a chosen starting address and a target reimbursement amount — a one-of-a-kind solution in France.
• Built with Symfony 2, MySQL (17M+ French addresses), ElasticSearch, OSRM...
• Implemented full-text search with autocomplete, high-performance routing engine, and tax/legal document generation.
• Installed monolithically on a dedicated server with hardened security (firewall, fail2ban, SSL...).
• Integrated Stripe payments, transactional emails with POSTFIX, and user management.
• Served thousands of customers with fast, reliable calculations and record-keeping.

Built a SaaS platform to manage queues and reservations for restaurants and public venues using Laravel, jQuery & Bootstrap.

• Enabled customers to join a virtual waiting line with their phone number.
• Provided businesses with tools to collect client contacts and send SMS notifications (opt-in).
• Integrated Stripe payments and a full SMS workflow.
• Fully developed and launched with a first client, but activity was discontinued due to lack of funding.

• First, I ran my own freelance business. I then founded 3 more companies in Toulouse (France), delivering IT and web solutions to 1,500+ clients (SMBs and individuals):
• Cestan Groupe Numérique (5 years): Web development (Symfony, Laravel, PHP...), FileMaker, and IT support. Built, among others, a large e-commerce platform with ERP integration for grocery delivery, and a social/dating website still active today with 50,000+ users.
• Handy (1 year, sold): Mac/PC, smartphone & tablet repair services.
• The Digital Smartness (2 years): Focused on advanced web technologies and FileMaker development (no more hardware).

• Airbus (Toulouse-Blagnac) — IT support & technician role, handling infrastructure and user assistance.
• Abalon (Communications agency) — Web developer (PHP), building and maintaining websites for clients.

Student working in IT alongside classes from age 13.