Tristan Faure

Data/AI Intern

Axian Group — MVola (Yas) · Antananarivo, Madagascar · Jul – Aug 2026

• Summer internship in the fintech arm of Axian Group — MVola, the leading mobile money service in Madagascar (Yas brand) — reporting directly to two group-level decision-makers (both ex-Orange France): the executive steering support quality of service, and the head of the group's target technical architecture.
• Designed and shipped to production "QoS Hub", a service-quality analysis platform for the group's three mobile money operators (Madagascar, Tanzania, Togo): 22,451 incidents unified from three heterogeneous sources, delivered as 15 pages of analysis. Python, FastAPI, PostgreSQL, Docker, Kubernetes.
• Industrialized an existing analysis chain into a multi-user platform: uploads through the UI, asynchronous worker processing, versioned and reversible publication — a failed import never touches the served version.
• Normalized three incompatible data formats: a 57-field ServiceNow export, a 15-field export in local time with durations in seconds, and a hand-maintained register with no identifiers. Idempotent converters, merge by ticket number, tolerance to naming variations.
• Authentication and isolation against Microsoft Entra ID: OIDC via oauth2-proxy, MFA, security-group filtering, per-country scoping enforced server-side. Three access layers validated by negative tests on the real environment.
• Deployed on a Kubernetes cluster: multi-architecture images, manifests, secrets, ingress, plus installing and operating PostgreSQL 16 on RHEL 9. Cross-team network diagnosis that identified an inoperative firewall rule and a reverse proxy limiting uploads.
• 129 automated tests and a full dry-run of a blank deployment that surfaced two defects before go-live.
• Documented handover: 15 published pages, from the recovery guide to the monthly operating procedure, plus full-rebuild procedures.
• Earlier analysis phase adopted by management: 39% false alerts identified as the #1 efficiency lever in Madagascar, and ~1/3 recurring incidents motivating formal problem management.

ML Research Project (PFEE) — Hodgkin Lymphoma

Hospices Civils de Lyon (HCL Sud) × EPITA · Lyon, France · 2026 – Present

• End-of-studies research project (PFEE) with the Hospices Civils de Lyon (HCL Sud hospital network): ML models for early detection and treatment-outcome prediction in Hodgkin lymphoma.
• Conducted under clinical data-governance constraints: anonymization, auditability, interpretability.
• Working toward a co-authored academic paper from this research.

Ecothon (EPITA × onepoint) — 🏆 1st Place, onepoint Prize

EPITA, in partnership with onepoint · Lyon, France · Jul 2026

• Won 1st place (onepoint Prize) at the sustainability hackathon organized by EPITA with onepoint (major French digital consulting firm, ~3,500 people) — team of 5.
• Built a cloud architecture carbon-impact simulator answering one question: which VM, at which provider, in which country, for the real workload need — at the best carbon and best price.
• Core insight: a CPU % is relative to the machine that carries it, so the engine converts measured load into absolute vCPU need before comparing instances — and surfaces the "FinOps trap" (e.g. Virginia is 14% cheaper but 3.8× dirtier than comparable zones).
• Radically sober engineering to match the message: 100% static page, zero npm dependencies, zero runtime network calls (versioned Boavizta + pricing data), build budget enforced at <100 KB gzip.
• Engine validated against the Boavizta API with a maximum deviation of 0.42% — every claim replayable via automated tests (npm run verify).
• Full deliverables beyond code: problem reframing, stakeholder mapping, sustainability & ethics briefs, and a timed 5-minute pitch.

World Model Hackathon (EB-JEPA) — Qualified Top 100 / 650 teams

Hackathon hosted by Yann LeCun · France · Jun 19–20, 2026

• Qualified in the top 100 out of 650 teams and competed on-site at the World Model Hackathon hosted by Yann LeCun (Meta AI Chief Scientist, Turing Award), themed on EB-JEPA — LeCun's proposed alternative to autoregressive LLMs, predicting in abstract representation space instead of pixels or tokens.
• Competed on two tracks, delivering a full research pipeline and 12–13 slide scientific presentation on each in 2 days.
• Clinical EEG track — "Physic Latent JEPA": self-supervised JEPA on the TUAB EEG corpus (2,717 recordings) with physics anchors (amplitude, phase, bandpower) tying the latent to physiological meaning; hypothesis: purely geometric regularizers (VICReg-style) prevent trivial collapse but allow functional collapse. Reached 79.0% balanced accuracy with a patient-disjoint probe, beating masked, temporal and random baselines, with the healthiest latent (effective rank 14.8, 0% dead dimensions).
• World-models track — "H-JEPA": a strict 2-level Hierarchical JEPA navigating 21×21 pixel mazes entirely in latent space, with no A* in the decision loop — an action-conditioned frozen world model, a coarse abstraction trained on latent "dreams", and a learned quasimetric as an A*-free compass.
• Key result: adding visit memory and a receding-horizon subgoal to the controller lifted success from 45% to 86% on 200 seeded mazes (organizer baseline ≈ 66%), with honest negative results documented (e.g. sharpening the metric with hard negatives regressed 86% → 75.5%).

Business Developer (DevCo)

JECT — Junior EPITA Conseil Technologies · Lyon, France · Feb 2026 – Present

• Active member of EPITA's Junior Enterprise as a Business Developer.
• Prospecting and developing tech missions for companies (web/mobile dev, data, automation, AI).
• Writing commercial proposals and managing client relationships.

Software Developer

Paris Corporate Housing · Paris, France · Sep 2025 – Feb 2026

• Automated internal processes for a corporate real estate company (~25 employees), within a 6-person IT team.
• Designed multi-system automation workflows (n8n, Jenkins, AWS, Salesforce), saving ~40h/week on internal operations.
• Key project: automatic property inspection comparison pipeline (PDF images → Jenkins preprocessing → OCR → structured extraction → Salesforce LWC validation → client email generation).
• Automated 50-100 daily management emails (backup, assignment, Salesforce tagging), integrated Signal into n8n via signal-cli.
• Security & compliance: least-privilege IAM, GDPR anonymization, S3 encryption, Vault, WireGuard, AWS Bedrock for compliant AI.
• Drove adoption among field agents (80-90% adoption rate). Documented everything in Confluence.

Founder

Decliik · Lyon, France · 2026 – Present

• Founded and building a SaaS platform that automates job search with AI — scraping, matching, CV/cover letter generation.
• Scrapes 4 major job boards (LinkedIn, Indeed, Welcome to the Jungle, HelloWork) daily with AI-powered compatibility scoring.
• B2B pivot targeting schools: admin dashboards, student tracking, analytics, and partner job offers.
• Currently in beta at decliik.io with early users and institutional prospects.

Co-Founder

AutoPilot.bg · Sofia, Bulgaria · 2026 – Present

• Co-founded a business process automation agency selling workflow automation to Bulgarian companies.
• Automating CRM, invoicing, lead generation, and admin processes — saving clients 15+ hours/week.
• Building automation workflows with n8n, Make.com, and custom API integrations, with AI (ChatGPT) integration.
• Serving industries including real estate, e-commerce, SaaS, recruitment, and accounting.

Co-Founder & CTO

Kanel Records → Kanel IT LLP · London, UK · Aug 2023 – Feb 2025

• Co-founded Kanel Records LLP at 19, initially building an AI music generation system creating virtual artists for Spotify.
• Pivoted the company to Kanel IT — a web & app development studio selling custom websites, web apps, and mobile applications to clients.
• Led a commission-based dev team, managed the full project lifecycle from prospecting to delivery: client meetings, quotes, contracts, development, and deployment.
• Scaled operations through business introducers, budgeting, pricing strategy, and partnership development.

Computer Engineering — SCIA-G (AI specialization)

EPITA · Paris → Lyon, France · 2021 – 2027

Baccalauréat — Mention Bien (15.7/20)

Lycée Français de Tananarive · Antananarivo, Madagascar · 2014 – 2021

EdgeOrchestra (2025–2026)

Federated self-supervised learning on Apple devices — a LeJEPA research testbed

CIFAR-10 CNN (2025)

Image classification with progressive CNN architectures

MNIST Classifier (2025)

99.69% accuracy handwritten digit recognition

OCR Sudoku Solver (2025)

Photo → digit recognition → solved puzzle