AI Research Intern

Meetdavis; Paris; Hybride; EUR 13 000 - 20 000; Temps partiel

Résumé du poste

Davis AI in Paris is seeking an AI Research Engineer intern to push the state-of-the-art in architectural-design diffusion models. You will collaborate with architects to turn foundational research into deployable tools and run falsifiable ablations on our discrete diffusion model.

You will join a focused team of engineers and researchers to improve training, conditioning, and sampling at multi-GPU scale, and help build the vision pipeline that converts raw plans into training data.

Qualifications

• Currently enrolled in a top MSc/PhD in ML, CS, applied maths, or a closely related field.
• Strong understanding of diffusion models (discrete or continuous) and generative modeling.
• Proficient in Python with PyTorch; able to write clean, scalable code.
• Experience with distributed training and modern ML tooling is a plus.

Responsabilités

• Experiment with model architecture and design space exploration to falsify hypotheses.
• Improve SOTA models with multi-GPU training, loss design, and sampling.
• Build the data pipeline by parsing plans and converting them into structured representations.

Connaissances

Applied Research Excellence

Diffusion Model Expertise

Python

PyTorch

Research Literacy

Collaboration & Communication

Formation

MSc/PhD in ML/CS

Outils

PyTorch Lightning

Distributed Training

Description du poste

TLDR: Davis AI is hiring an AI Research Intern to push further and scale our state-of-the-art floorplan generation model. Image-based models produce plans that break building codes under scrutiny; ours works on structure, not pixels, and is already beating all the benchmarks. You will be joining the tech team and, alongside architects, you will run falsifiable ablations on our discrete diffusion model, improve training, conditioning, and sampling at multi-GPU scale, and build the vision pipeline that turns raw plans into training data.

About Davis

Davis is an AI-native real estate company accelerating early-stage development and architectural design. Today developers coordinate 4-5 fragmented stakeholders over weeks or months. Soon they'll need only one: Davis.

We turn every input that shapes a development decision into decision-ready outputs: investor-grade feasibility studies, investment analysis, and architect-certified designs, delivered in days. Every stage pairs our proprietary AI systems with expert review, so velocity never comes at the cost of reliability.

We closed a $5.5M pre-seed co-led by Heartcore Capital and Balderton Capital, with Yellow, Evantic and Entrepreneur First, alongside angels from the founding teams of Spacemaker, Black Forest Labs, Hugging Face, Supabase, Cleo and Spore Bio. We already work with leading developers and expect to support hundreds of projects over the coming year, deepening our research, our hiring, and our coverage of the development process end to end.

Our Mission

We build a foundation model for architectural design that generates compliant, editable building layouts from scratch. By leveraging discrete diffusion models (operating on structured representations rather than pixels), we aim to produce floorplans and site plans that respect real-world constraints (zoning laws, space requirements, etc.) and can be iteratively refined like a human-designed plan.

The Role

We are looking for an AI Research Engineer intern to spearhead this effort in our Paris office. If you’re excited about pushing the state-of-the-art in generative models and applying it to a high-impact domain, this role offers a unique opportunity to define a new class of AI-driven design tools. You will work within a focused team of 3 engineers and researchers, collaborating daily with architecture experts to turn foundational research into deployable tools.

What you’ll be working on:

• Model Architecture & Design Space exploration: Take an open question about the model and answer it properly. Read the literature, form a hypothesis, design the ablation that can actually falsify it, run it, and make sense of what you found.

• Model improvement: Improve our SOTA model. That means training runs on multi-GPU nodes, reading loss curves and generated samples, and iterating on architecture, loss design, conditioning, and sampling.

• Dataset creation: Training data is the single biggest lever on output quality. The work is building the agentic pipeline that produces it automatically, a computer vision problem at its core: parsing raster plans, PDFs and scans into a clean, consistent structured representation.

What We’re Looking For

• Applied Research Excellence: Currently in a top MSc/PhD (or equivalent) in ML, CS, applied maths, or a closely related field.

• Diffusion Model Expertise: Strong understanding of diffusion models (discrete or continuous), guided generation techniques, and the latest advances in generative modeling.

• Technical Engineering Skills: Strong programmer in Python with experience in PyTorch. Ability to write efficient, maintainable code and optimize training pipelines. Familiarity with distributed training libraries (e.g., PyTorch Lightning) is a plus.

• Research Literacy: Ability to read, evaluate, and implement advanced ML research. You stay current with state-of-the-art generative modeling work and can adapt cutting‑edge methodologies to domain‑specific problems.

• Collaboration & Communication: Strong teamwork skills. Able to articulate complex concepts clearly and work closely with AI engineers, researchers, and domain experts to integrate technical solutions into the architectural design workflow.

Nice to Have

• Experience : Former internship in Diffusion or generative modelling

• Publications/Open Source: Research publications in generative modeling (diffusion, VAEs, flows, GANs) or significant open‑source contributions demonstrating ability to push state-of-the‑art systems.

• Domain Knowledge: While not required, an interest in architecture or design will help.

• Reinforcement Learning & Optimization: Experience with RL, reward modeling, or constrained optimization (particularly MCTS, GRPO and RLHF) relevant to guiding generative models under complex constraints.

• Multimodal Generation: Experience with graph-based, sequence-based, or discrete structured generative models (e.g., molecule generation, layout generation, program synthesis) or with graph neural networks.

• Computer Vision: Experience with document and structured‑image understanding, such as VLMs, OCR and layout parsing, vectorization, or segmentation and detection pipelines applied to plans, diagrams or CAD‑like sources.

• Agentic Systems: Experience building LLM agent pipelines — tool use, multi‑step workflows, prompt optimization frameworks (DSPy, GEPA) — and the evaluation harnesses needed to make them reliable at scale.

Why Join Us

You're joining a team of 12 at the very beginning - where every decision you make shapes the product, the culture, and the trajectory of the company. What you build here will be yours.

• Build what no one has before: the foundation model that automates architectural design and redefines how cities are imagined, designed, and built.

• Work on meaningful challenges: from constraint‑aware generative models to real‑world deployment in major construction projects.

• Competitive salary and meaningful equity in an early‑stage company.

• Ship fast, iterate boldly: go from research to prototype to production in weeks, not years.

• Join a world‑class team: a mix of AI researchers, engineers, and architects backed by world‑class VCs.

More information about Davis, the team and the market we’re going after:

Team

Mehdi (Co‑founder & CEO) grew up in a family of architects and has lived this problem firsthand. He's a repeat founder who bootstrapped his first startup at 20, and graduated from Sciences Po and HEC Paris. Amine (Co‑founder & CTO) is an AI researcher from École Polytechnique who worked extensively on discrete diffusion and turned down a PhD with Google DeepMind to build Davis. They started working together in July 2025 at Entrepreneur First’s first European residency, a two‑month lock‑in in a German castle.

Today we’re a team of 12: technical profiles from Polytechnique, ENS and INRIA alongside architects and deep real estate expertise.

We’re small with an extremely high bar. If you want to work deeply on hard problems and see your work reach clients within days, you’re the one we need.

Why we'll win

Real estate is a $13 trillion industry that technology has largely bypassed. The professional services that feed it (design, engineering, feasibility, permitting) represent hundreds of billions in spend that no one has seriously automated.

Proptech spent the last decade selling SaaS on the edges of these workflows. It didn’t work, for two reasons: no professional wants another tool to learn, and no tool can automate work that runs on expert judgment. Davis makes a different bet. We don’t sell tools, we sell the work: AI-generated, expert‑validated, delivered in days instead of weeks. Every project compounds our data advantage across typologies, geographies and regulatory contexts.

Why No One Has Solved Architectural Design

Real estate development bleeds time and money in architectural design loops. Architects cycle through dozens of floorplan revisions to meet regulatory and client constraints, each round taking days, each missed constraint restarting the loop. Traditional CAD and BIM tools offer zero generative capability; parametric tools only check constraints after generation, leading to designs that frequently break under new zoning rules or irregular sites.

Generative AI has the potential to solve this, but doesn’t yet. Fine‑tuning image diffusion models on floorplans produces layouts that look plausible but fall apart under scrutiny: hallucinated rooms, mislabeled spaces, code violations no architect would accept. Pixel‑space models have no concept of what a wall is or why corridor needs to connect two things. Compliance‑guidance techniques typically required segmentation at each noisy timestep, compounding errors and making major edits impractical. As a result, floorplans may ’look’ plausible but break building codes, or need heavy post‑processing before they’re usable.

We care about who you are, not just what’s on your CV.

If you're drawn to what we're building but don't meet every requirement, we still want to hear from you. Studies show that women in particular tend to apply only when they meet 100% of the criteria. If that's you, please don't let that hold you back. We'd love to receive your application.