# TW3 Partners

> TW3 Partners builds AI agents that work inside its clients' business applications and draft their deliverables, deployed on the cloud chosen by the client, on its servers or in an environment without any outside connection. Each answer cites the document, the page and the paragraph it is drawn from.

French version: https://tw3partners.fr/

TW3 Partners is a Paris-based AI engineering and research firm. The team has 15 engineers and PhDs, all employees, with no subcontracting. It has two activities: engineers who build AI systems inside client teams, from scoping to production, and research, carried out by its subsidiary Racine.ai.

What the agents do. Each answer cites the document and the page it comes from. When the information is not in the client's documents, the agent says so instead of filling the gap. Every request is logged in the client's environment. Agents connect to business software (Salesforce, SAP, Moodle, Microsoft 365, SharePoint, HubSpot) through APIs or the open MCP standard, and through the screen when there is no API.

Hosting and models, chosen by the client: Azure, AWS or Google Cloud with the models the client already uses; European hosting (Scaleway, OVHcloud) with European open models; the client's own servers, including air-gapped environments. Models from Anthropic, Google, Mistral AI, GLM, Qwen, or the open models of Racine.ai.

How an engagement works. Work starts with one use case and a success criterion set at the start; the client's teams judge the result on their own documents before any wider commitment. Code and documentation are delivered in the client's repositories, so the client can take the system over. Billing is time and materials or a fixed price per work package. TW3 Partners resells no licence and no cloud capacity. No prices are published. A client is named only with written authorization.

## Proof points

- TW3Cast, 3rd out of 130 on the GIFT-Eval forecasting benchmark (mean MASE rank). Source: public GIFT-Eval leaderboard, 14 September 2026. Paper: arXiv 2609.28506, https://arxiv.org/abs/2609.28506 . Code: https://github.com/TW3-Partners-OS/TW3-Cast
- CU-1 (released as UI-DETR-1): 70.8% successful clicks on the WebClick benchmark, against 58.8% for Microsoft's OmniParser V2, both at optimal settings. Source: https://huggingface.co/racineai/UI-DETR-1 . Paper: IEEE Big Data 2025, https://doi.org/10.1109/BigData66926.2025.11402585
- Open models and datasets downloaded over 1.4 million times on Hugging Face (Hugging Face API, 19 September 2026): https://huggingface.co/racineai
- Aalto University's DistilVDR trains on 454,000 images from our open VDR datasets. Source: arXiv 2608.10636, https://arxiv.org/abs/2608.10636
- Content scoring for AI engines: https://arxiv.org/abs/2609.07559
- "a very agile start-up at the forefront of AI technologies": Marc Blangy, Group Chief Information & Transformation Officer, OMNES Education, IT for Business no. 2317, June 2026, p. 18 (translated from French).
- Racine.ai holds the Bpifrance Deeptech label and received the French Tech Émergence grant. TW3 Partners holds the Jeune Entreprise Innovante status, is listed in the Osez l'IA catalogue of the French Directorate General for Enterprise, and has been a member of France Deeptech since July 2025. Citead, its GEO platform, won the VivaTech Challenge Région Sud in June 2026.

## Services

- [Forward Deployed Engineers](https://tw3partners.fr/en/services/forward-deployed-engineers): Our engineers join your teams, work in your tools and stay through to production.
- [AI & Cloud Strategy](https://tw3partners.fr/en/services/strategie-ia-cloud): We don't resell licences or cloud capacity.
- [AI in your applications](https://tw3partners.fr/en/services/ia-applications-metier): The agent works in Salesforce, SAP, Moodle, Microsoft 365, SharePoint or HubSpot, with the permissions of the person using it.
- [Agentic platforms](https://tw3partners.fr/en/services/plateformes-agentiques): Our agents draw on your document collections to draft reports, tender responses and regulatory filings, even from very large volumes such as decades of technical reports.
- [GEO (visibility in AI engines)](https://tw3partners.fr/en/services/geo): Your buyers ask ChatGPT, Claude, Gemini and Perplexity before they contact you. Citead, our GEO (Generative Engine Optimization) platform, measures which sources those answers cite, and our teams rewrite your pages so that they are cited too.
- [Training & maintenance](https://tw3partners.fr/en/services/formation-maintenance): After go-live, your teams learn to use the systems we deliver, and we keep models and connectors up to date.

## Sectors

- [Energy](https://tw3partners.fr/en/secteurs/energie): Our agents make your reports searchable and prepare your technical proposals; you know your solar power generation the day before.
- [Pharma and chemicals](https://tw3partners.fr/en/secteurs/pharma-chimie): Our agents prepare a first draft of your regulatory filings and read diagrams, tables and handwritten documents, on a European cloud or on your own servers.
- [Industry, construction and engineering](https://tw3partners.fr/en/secteurs/industrie-batiment): Our agents make your procedures searchable, prepare your tender responses and consolidate room occupancy schedules across your sites.
- [Higher education](https://tw3partners.fr/en/secteurs/enseignement-superieur): Our agents prepare your accreditation files, enter data into your software, without modifying it, and answer questions using your internal content.

## Use cases

- [Use cases](https://tw3partners.fr/en/references): Use cases described in general terms, with the tools used. A client is named only when they have authorized it in writing.

## R&D

- [Research programme](https://tw3partners.fr/en/r-d): We train our own models, publish our results and open our datasets. More than 1.4 million downloads on Hugging Face (September 2026).
- [Publications](https://tw3partners.fr/en/r-d#publications): Each publication with its link: arXiv, IEEE Big Data 2025, ECML PKDD 2026, the Hugging Face blog.
- [Models and datasets](https://tw3partners.fr/en/r-d#modeles): Each model and dataset with its Hugging Face link and its downloads.
- [Racine.ai](https://tw3partners.fr/en/r-d#racine-ai): The R&D subsidiary of TW3 Partners, holding the Bpifrance Deeptech label.

## Company

- [About](https://tw3partners.fr/en/a-propos): The team of 15 engineers and PhDs, labels and awards, scientific partnerships, and the group structure with the subsidiary Racine.ai.
- [Contact](https://tw3partners.fr/en/contact): Contact form: a written reply within one business day, then a one-hour discussion with the technical teams and a partner.

## Questions

### How soon do we see results?

Within a few weeks. We start with one use case and a success criterion set at the outset. Your teams judge the result on their own documents before any wider commitment.

### Can you work with sensitive data?

Yes. Your data goes only where you decide: your cloud, a European cloud, your servers, or an air-gapped environment. Our engineers work only with the access you grant them. When data must not leave, we use open models hosted in your environment.

### How do you check the AI's answers?

Every answer cites the document and page it comes from, so your teams can open the source and check. When the information is missing, the agent says so.

### How do you bill?

Our engagements are billed on a time-and-materials basis or at a fixed price per work package.

### Why choose you over a large systems integrator?

The engineer you meet first is the one who builds your system. Code and documentation are delivered in your repositories, so your teams can take over the system without us.

## Optional

- [Press](https://tw3partners.fr/en/presse): Press coverage, interviews and public talks by the TW3 Partners team.
- [Blog](https://tw3partners.fr/en/blog): Articles published by TW3 Partners (in French).
- [Careers](https://tw3partners.fr/en/carrieres): Working at TW3 Partners: teams, culture, sport, hiring process, and open positions at TW3 Partners and Racine.ai in full.