Job Description

Location: Paris (on-site)

Contract: Freelance

Duration: 3 months · 4 days/week

Start: ASAP

Languages: French and English (fluent in both)

About Colonies

Colonies is on a mission to solve housing. We build and manage flexible living spaces across Europe. The tech team is small and high-leverage. We run a hub-and-spoke model: a Core Tech team maintains the platform, while embedded engineers position themselves in business teams to solve problems close to the source. What works gets hardened and scaled. What doesn’t gets killed quickly.

The job

You will build AI-powered automation for operational workflows that are currently manual and slow. The first mission is inventory processing: tenants move out, inventories are completed (photos, comments, documents), and today a human reviews every one to assess damage, estimate costs, and decide what to deduct from the deposit. You will automate that end-to-end.

Concretely, expect to:

• Ingest inventory documents — photos, written comments, structured reports — and extract structured damage assessments using vision and language models

• Compare move-out inventories against move-in inventories to identify new damage and attribute responsibility

• Estimate repair costs based on damage type, severity, and historical data

• Feed results into the deposit return process so tenants get their money back faster and ops teams stop reviewing every file by hand

• Build this as a reusable document processing pipeline — the same underlying tooling (ingestion, extraction, comparison, decision logic) will apply to other workflows after inventories

The stack will likely involve LlamaIndex, LangChain, or equivalent frameworks for orchestrating LLM-based pipelines, alongside Python, our internal APIs, and Postgres. You will collaborate closely with our other freelance engineers — particularly the analytics engineer working on data infrastructure — and report into the Core Tech team for standards and architecture.

Once the inventory pipeline is live, the role extends to applying the same document processing and AI capabilities to other high-impact operational workflows.

What this is not

This is not a research role. We are not looking for someone to evaluate models and write benchmarks. We need someone who can take a messy real-world process, design an AI pipeline that handles it reliably, ship it into production, and iterate until it works well enough that the ops team trusts it. You own the problem from raw document to automated decision.

You

• You have built LLM-powered applications in production — not just prototypes. You know how to handle hallucinations, edge cases, and the gap between demo and reliable system.

• You are strong in Python and comfortable working with frameworks like LlamaIndex, LangChain, or equivalent orchestration tools for building retrieval, extraction, and reasoning pipelines.

• You have experience with vision models and multimodal AI — extracting structured information from images and documents is core to this role.

• You understand how to design evaluation and feedback loops for AI systems: measuring accuracy, catching regressions, and improving over time.

• You can work with APIs, databases (Postgres), and existing internal systems. You do not need a clean abstraction layer handed to you.

• You use AI tools (Copilot, Claude, Cursor, etc.) as part of your daily workflow, not as a novelty.

• You have worked in fast-moving environments where you had to figure out both the problem and the solution yourself.

• You can talk to non-technical operations teams, understand their process and edge cases, and translate that into system requirements.

• You are comfortable with ambiguity and have a bias for action. The current process is manual and poorly documented. You will need to reverse-engineer it before you can automate it.

• You have a proven track record. You can point to AI-powered systems you have built that run in production and handle real workloads.