# Alexis Thomas

> Applied Scientist at Amazon building demand forecasting systems for European logistics, and an engineer who ships AI products end to end.

I'm an Applied Scientist at Amazon in London, where I build the demand forecasts that Amazon uses to plan its European delivery network, from three-year capacity plans down to zip-code-level operations. Before moving into science I spent two and a half years as a software engineer at Amazon, building the serverless platforms that run our research models in production.

- Current role: Applied Scientist II at Amazon (London, United Kingdom)
- Email: contact@alexisthomas.fr
- LinkedIn: https://www.linkedin.com/in/alexis-thomas11/
- Google Scholar: https://scholar.google.com/citations?user=pdpHNQUAAAAJ
- GitHub: https://github.com/alexis-thomas

## Pages

- [Résumé](https://alexisthomas.fr/resume): full work history, education, skills and certifications
- [Projects](https://alexisthomas.fr/projects): case studies of apps I built
- [Prismo](https://alexisthomas.fr/projects/prismo): An AI learning journal that turns what you read, watch and hear into organized notes and spaced-repetition quizzes.
- [Giftruly](https://alexisthomas.fr/projects/giftruly): An AI gift concierge: describe the person in a sentence and get a shortlist of ideas, each with a reason it fits.
- [Aura](https://alexisthomas.fr/projects/aura): A private mood journal with an AI companion: check in with a tap, and get a gentle note back.
- [Research](https://alexisthomas.fr/research): publications and awards
- [CV (PDF)](https://alexisthomas.fr/Alexis-Thomas-Resume.pdf)
- [Contact](https://alexisthomas.fr/contact)

## Experience

### Applied Scientist II, Amazon (Oct 2024 – Present, London, UK)

- Built a long-term demand forecasting system for a 3-year planning horizon, combining optimization algorithms with statistical modeling to satisfy multiple business constraints across 10 marketplaces and 31 countries, reaching 99.7% constraint satisfaction.
- Developed a short-term granular forecasting model for 2-month operational horizons, using hierarchical neural networks and statistical time series models to predict demand at zip-code level with an average WAPE of about 3%.
- Led a model evaluation programme across statistical, machine learning and time series foundation models, and designed adaptive forecasting strategies that adjust automatically to different volatility regimes and market conditions.

### Software Development Engineer II, Amazon (Dec 2023 – Sep 2024, Paris, France)

- Owned architecture, development and engineering standards for a science team optimizing the Amazon delivery network in Europe.
- Engineered long-term demand forecasting pipelines with machine learning and statistical models at granular levels.
- Implemented CI/CD pipelines with containerization to deploy operational research models on scalable serverless architectures.

### Software Development Engineer, Amazon (Mar 2022 – Dec 2023, Paris, France)

- Built a platform for one-click execution of operational research workflows, scaling cost-effectively to thousands of concurrent executions with no pre-provisioned servers.
- Delivered input validation, sanitization and infrastructure-as-code tooling that cut the time to put models in production and made them easier to maintain.
- Built a data visualization platform used by hundreds of internal users, with automated reporting, data pipelines and big data processing.

### Finance Project Manager, BMW Group France (Feb 2021 – Jul 2021, Paris, France)

- Worked with project managers to frame projects and track them to completion.
- Produced financial and staffing estimates and identified risks and opportunities across the BMW France project portfolio.

### Software Development Engineer Intern, Amazon Web Services (Jun 2020 – Dec 2020, Dublin, Ireland)

- Built a new service that lets customers adjust internal tool configurations in real time.
- Owned the project end to end: scoping, design, implementation and testing.
- Java and AWS on the backend, React on the frontend.

## Projects

### Prismo (2026 – present, Live on iOS & Android)

An AI learning journal that turns what you read, watch and hear into organized notes and spaced-repetition quizzes.

- A multi-stage LLM pipeline with deterministic guardrails: Entries go through understand → write → validate → repair. The model works from a compact summary of the user's library rather than the whole thing, and a deterministic, unit-tested validator rejects unsafe changes (duplicate or cyclic folders, deleting notes it hasn't read) before anything reaches the user, with at most one repair call. An offline evaluation set tracks quality against a cost target of under $0.50 per user per month.
- Offline-first sync across devices: Libraries sync through S3 with vector clocks for conflict detection and tombstones for deletions, so edits made offline on two devices merge without losing data.
- Stack: React Native (Expo), TypeScript, Swift (iOS widget), AWS Lambda, API Gateway, DynamoDB, S3, Cognito, AWS CDK, Gemini, Claude on Amazon Bedrock
- Links: https://prismo-app.com, https://apps.apple.com/app/id6759450471, https://play.google.com/store/apps/details?id=com.elial.prismo

### Giftruly (2023 – present, Live on web, iOS & Android)

An AI gift concierge: describe the person in a sentence and get a shortlist of ideas, each with a reason it fits.

- Automatic kill switches for LLM spend: CloudWatch alarms on the chat function trigger Lambdas that either disable the AI feature gracefully or throttle its concurrency, so a traffic spike or abuse can't turn into a surprise bill.
- A recommender that degrades gracefully: Giftruly 2.0 ranks a gift catalog locally, adds semantic search with embeddings, and asks Claude on Bedrock for the final picks with structured output. Model results go through the same validation as local ones, and spend limits are reserved atomically before each call; when a limit trips, users get local results instead of an error.
- Migrations without breaking shipped apps: The AI layer moved from OpenAI to Azure OpenAI to Claude on Bedrock. When Amazon retired its product API and an old Claude model reached end of life, I patched the production backend in place so the apps already on people's phones kept working without a new release.
- Stack: Next.js, React Native (Expo), TypeScript, AWS Lambda, DynamoDB, CloudFront, Cognito, AWS CDK, Claude on Amazon Bedrock, Embeddings
- Links: https://giftruly.com, https://apps.apple.com/us/app/giftruly-ai-gift-finder/id6463821360, https://play.google.com/store/apps/details?id=com.elial.Giftruly

### Aura (2024 – present, Live on the App Store)

A private mood journal with an AI companion: check in with a tap, and get a gentle note back.

- Serverless AI backend with cost protection: The app talks to a small AWS backend (API Gateway, Python Lambdas, Claude on Bedrock) without requiring an account: requests are signed with anonymous Cognito credentials. If usage spikes, a CloudWatch alarm alerts me and automatically throttles the AI function.
- In progress: a privacy-first rebuild: The next version keeps journal entries on the phone, runs on a backend with no database that never logs what people write, and falls back to on-device behaviour when the AI is unavailable.
- Stack: React Native (Expo), TypeScript, AWS Lambda (Python), API Gateway, Cognito, AWS CDK, Claude on Amazon Bedrock
- Links: https://apps.apple.com/us/app/aura-daily-journal-mood/id6480318800

## Education

- Mines Paris – PSL: Diplôme d'ingénieur civil des Mines (Master's degree), Mathematics & Computer Science (2018–2022)
- Lycée Saint-Louis: Classes préparatoires aux grandes écoles (CPGE) (2016–2018)

## Publications, awards & certifications

- Forecasting Electric Vehicle Charging Station Occupancy: Smarter Mobility Data Challenge. Journal of Data-centric Machine Learning Research (DMLR), 2024. https://data.mlr.press/assets/pdf/v01-16.pdf
- 3rd place, Smarter Mobility Data Challenge (2023-03)
- AWS Certified Solutions Architect – Professional (2024-05) https://www.credly.com/badges/ec115f20-7880-4b9d-bdcc-bdb29e22bd5a

## Skills

- Forecasting & ML: Time series forecasting, Hierarchical forecasting, Neural networks, Foundation models for time series, Statistical modeling, Optimization, Model evaluation
- AI engineering: LLM applications, Amazon Bedrock, Embeddings & semantic search
- Cloud & infrastructure: AWS (Lambda, ECS/Fargate, SageMaker, Glue, Redshift, DynamoDB, S3, CloudFront), AWS CDK, Docker, CI/CD pipelines, Serverless architecture, Observability
- Languages & frameworks: Python, TypeScript, Java, SQL, React, React Native / Expo, Next.js, Spark