Independent engineering studio · Paris, France

Mobile products that think. AI agents that work.

We design Android and cross-platform applications, and the agent systems that make them autonomous: understanding, reasoning, voice, decisions. Twelve years of mission-critical work for major brands, now applied to AI in production.

Start a project Our expertise

12 yrsof production software engineering

2M+users on a single application

18countries in one deployment

~99%stability on a critical app

agent session · mobile-zillustrative scenario

mission: give the product a voice assistant ▸ execution plan 1. audit the existing mobile architecture 2. draw the device ↔ cloud boundary 3. prototype the voice loop 4. measure latency, cost, privacy ▸ audio capture · on-device ▸ local transcription · 23 ms · never sent ▸ reasoning · cloud ▸ Claude · tool use · streaming ✓ full voice loop · 480 ms perceived ✓ cost / session −62% vs full-cloud ✓ private data never leaves the device // illustrative scenario: indicative figures

edge · LiteRT-LMcloud · Claude APIloop · MCP tools

missions delivered forLouis VuittonThalesFrance TélévisionsRenaultDeezerRATPTotalEnergiesYaraL'OréalCrédit du Nord

expertise

The device ↔ cloud boundary,
drawn to the millimetre.

The real decision in an AI product is not about prompts. It is about what runs on the device and what goes to the model, in latency, cost and privacy. That is our specialty: knowing where to draw the line, and building both sides.

on-device

What stays on the device

Private, instant, offline. The part of the system users should never have to wait for, or send.

[a]Local voice capture & transcription : the microphone never leaves the pocket
[b]Embedded small models : LiteRT-LM, ExecuTorch, quantization
[c]Cache, index & embeddings : the product's memory, on-device
[d]Moderation & guardrails : first filters before any network call

cloud · reasoning

What goes to the model

Heavy reasoning, tools, world knowledge. Multi-provider architectures including Claude, in production.

[e]Agents & tool use : execution loops, MCP, tool calls
[f]Multi-model orchestration : Claude, open-weight models, cost/quality trade-offs
[g]Streaming & real-time : voice, turn-taking, perceived latency
[h]LLM telemetry : cost, tokens, quality: every decision measured

local latency · no network round tripdistributed reasoning · streaming

product engineering

And underneath both: an industrial-grade mobile base.

Native Android Kotlin, Kotlin Multiplatform shared across Android, iOS and web, Kotlin back-end services (Spring Boot, Ktor), real-time IoT (MQTT, AWS IoT Core), modular architecture, security (OWASP, OAuth 2.0 + PKCE), full CI/CD. At Louis Vuitton, moving to a multi-module architecture cut build time by 40%. The ground agents stand on, because an agent on a fragile base is not a product.

work

A system that speaks is a system that runs.

Two representative engagements: Iris, a product in production co-founded by Houssem Zaier, and Code Companion at Thales, in a secured environment. Both with verifiable numbers, not screenshots.

Iris · consumer conversational & voice AI product

in productionco-founded by Houssem Zaier

A consumer conversational and voice product powered by several LLMs, Claude among them. Houssem's scope: he founded the Kotlin Multiplatform module shared by Android, iOS and the web, with a Compose UI that runs all the way to the browser through Kotlin/Wasm, and built the real-time voice layer on all three platforms: microphone capture, streaming speech recognition and synthesis, turn-taking. He also prototyped a “local brain”: a small quantized LLM running on the phone answers instantly while the cloud reply is on its way.

1,200+commits in 6 months

3platforms, one Kotlin codebase

Local + cloudon-device LLM prototype

Real-timevoice and turn-taking

Kotlin Multiplatform · Compose Multiplatform · Kotlin/Wasm · Swift · real-time voice over WebSocket · llama.cpp · render-test CI

Thales · Code Companion, the internal AI coding assistant

secured environment2 PRs merged upstream

Nine months of full-stack work on Code Companion, the AI coding assistant used by the group's developers, built on a fork of the open-source Continue: IntelliJ plugin in Kotlin, VS Code extension, shared TypeScript/Node.js core, React interface, Python service orchestrating LLM calls (multi-provider routing), and agentic features (tool calls, multi-step execution). Regular rebases from upstream, with the security-specific code kept isolated. Two improvements contributed back and merged: PR #8524 (IntelliJ plugin memory allocations) and PR #9659 (parallel tool calls).

9 monthsfull-stack engagement

2 PRsmerged into Continue

2IDEs covered: IntelliJ + VS Code

Multi-LLMrouting and agentic features

Kotlin (IntelliJ SDK) · TypeScript · Node.js · React · Vite · Python · LLM tooling · IPC

12 yrsof production software

2M+users served

~99%stability, critical e-commerce app

Since 2024agents & MCP in daily practice

the studio

A practice of expertise, not a factory.

An independent studio, founded and led by Houssem Zaier.

Houssem Zaier

Founder · Principal Architect

Kotlin for Java Developers · JetBrains (Coursera) 2019PSM & PSPO · Scrum.orgFrench · English · Arabic

✓12 years of mobile technical leadership across luxury, media, transport, automotive, energy and defense, including applications serving millions of users.
✓Agentic practice since 2024 : from the day Anthropic opened the MCP protocol: agents, tool use, tool servers, daily and in production.
✓Co-founder of Iris: a voice AI product in production, seen from the inside, from the mobile client to model trade-offs.
✓A simple thesis: knowing what runs on the device and what goes to Claude, and being able to build both sides of the boundary.
✓A French company: SAS founded in 2022, EU VAT verified, SYNTEC collective agreement.

The studio works directly with engineering and product leaders, or alongside your usual partners.

contact

An AI product to ship?

Describe the context, even roughly. You will get a reasoned technical answer within 48 hours, and a free first device ↔ cloud arbitrage opinion.

✉ [email protected]☎ +33 7 81 49 09 88 in LinkedIn

Asnières-sur-Seine · Paris · Europe · on-site, hybrid or remote engagements