Forward Deployed AI Engineer

I build autonomous agents.

AI Engineer specialising in LLMs, VLMs, Model Context Protocol and fine-tuning. Currently Forward Deployed AI Engineer at Liberty International.

Explore my work

Scroll to explore

Hasan Basbunar / Forward Deployed AILuxembourg · 2026

Luxembourg

Hasan Basbunar

Forward Deployed AI Engineer

Forward Deployed · Luxembourg

About / 01

Where models meet production.

Forward Deployed AI Engineer at Liberty International, I'm embedded close to the business inside the group's companies, from pharma to hospitality, media and equestrian. My mission: automate redundant work and ship AI agents that take action.

Trained in Metz (BUT Data Science, top of class) then at UPEC (Master OIVM, distributed systems & data science), I push my craft through personal R&D: from the full fine-tuning of a 24-billion-parameter multimodal LLM to patches accepted into the upstream Linux kernel.

At a glance

Based in: Luxembourg · Grand Est, France
Status: Forward Deployed · Liberty Intl
Languages: French
Native
Turkish
Native
English
Professional working proficiency
German
A2 – B1
Interests: Tech & SOTA monitoringMusicReadingSkiing

In the wild / 02

The agent, alive.

A glimpse of my daily practice: plans, thinking cycles, file reads, MCP calls. Scroll to watch execution unfold.

Session · ~/hasan

hasan-basbunar-20260420ctx: 22% of 200kmcp: 4 toolsenv: prod

>Introduce yourself in 30 seconds.

Four years, five grounds, one obsession: agents that take action in production.

Bash(grep -r "production" ~/journey/)

└liberty · cleva · foyer · lem3 · lodos · talkink

Read 5 files

└liberty/agents-that-act.md · cleva/mcp.md · lodos-24b/README.md · talkink/SHIP.md · kernel/patches.md

Searched for "upstream"

└2 patches merged into Linux · page_pool / netdev · modpost / kbuild

Thinking

└How to compress a Forward Deployed role, two kernel patches and a shipped app into one paragraph?

Write(hero/headline.tsx)

└"I build autonomous agents that take action"

Bash(cat ~/.hasan/languages.json)

└français & türkçe · native · english B2 · deutsch A2 – B1

Profile · Hasan Basbunar · Forward Deployed AI Engineer

✓Liberty International · Forward Deployed AI · agents that act
✓Cleva · AI Tech Lead · MCP, VLMs, agents in prod
✓Lodos-24B · full fine-tune 24B · 20M TR tokens
✓Talkink · 100% on-device macOS dictation · shipped + Homebrew
✓Linux kernel · 2 patches merged upstream
Next chapter · building on my own terms

›

bypass permissions on (shift+tab to cycle)

Experience / 02

Four years of compounding impact.

From deep learning research to autonomous agents deployed across companies.

05.2026 – PresentLiberty International SAMulti-sector holding · Luxembourg

Forward Deployed AI Engineer · Work-study

Embedded inside the group's companies to automate redundant work and ship AI agents that take action.

Five grounds, one engineer: H2 Pharma, Vengoh Equestrian, SixSeven Restaurant, Premium Magazine, Premium Private Club.
End-to-end automation: from manual processes to tool-using autonomous agents (MCP, tool-use).
Agents that act on real business systems, not chatbots.

AI AgentsMCPPythonLLMsVLMsRAGAutomationTool Use

09.2024 – 04.2026ClevaInsurance Software · Paris

AI Engineer & Tech Lead · Work-study

Architecture and deployment of autonomous AI agents for complex business process automation.

Model Context Protocol (MCP): implementing interoperability standards for agentic workflows.
Production engineering: modern Python 3.10+, AsyncIO for reactive UIs, defensive programming.
VLM-based extraction: fine-tuning Vision Language Models to extract structured data from document images.

Python 3.10+AsyncIOMCPLangChainMistral AIOpenRouterGradioDockerHugging FaceVLMs

02.2024 – 08.20246moFoyer AssurancesInsurance · Luxembourg

Excellence Award

Runner-up · SD-SIF Best Data Science Internship Prize 2025

Data Scientist · Internship

Development and fine-tuning of AI models (VLMs) for document-processing automation.

End-to-end multimodal fine-tuning pipeline on insurance documents.
Labelling, evaluation, and production integration of trained models.
Exploration of speech-to-text on conversational channels.

PythonPandasPyTorchVLMsLabellingFine-tuningSpeech-to-Text

04.2022 – 06.202210wksLEM3CNRS · Research · Metz

Machine Learning Developer · Internship

Developed deep learning models (CNN, U-Net) for material image reconstruction.

Applied research under the supervision of a CNRS team.
End-to-end training and evaluation pipelines.

PythonPyTorchCNNU-NetDeep Learning

Open source · Upstream

Code in the Linux kernel.

Four patch series sent to the kernel, two merged into mainline. Low-level debugging, Fixes-tagged patches, reviewed and tested by maintainers.

net / page_poolMerged in mainline

page_pool · memory-provider leak on error path

Kernel networking core. Restores strict init/uninit symmetry in page_pool_create_percpu(): resource leak eliminated.

Fixes:0f9214046893View commit

kbuild / modpostMerged in mainline

modpost · stack buffer overflow

In do_input_entry() and do_dmi_entry(). Reviewed-by & Tested-by: Randy Dunlap.

Fixes:1d8f430c15b3View commit

scripts / sorttableSubmitted upstream

scripts/sorttable · off-by-one

Build-time sort regression. One-line surgical fix, proven by the sibling loop in the same file.

Fixes:200d015e73b4View thread

bpf / bpf_dbgSubmitted upstream

bpf_dbg · dual off-by-one

BPF debug tool (cmd_select, pcap_next_pkt). Dormant bug since 2013, verified by a deterministic PoC.

Fixes:fd981e3c321aView thread

All my patches on lore.kernel.org

Projects / 03

R&D, hackathons, and shipped tools.

Public experiments that push my technical limits.

R&D Project2025

Lodos-24B · Full fine-tuning of a 24B multimodal LLM

Complete fine-tuning of Mistral Small 3.2 (24B parameters) on a 20M+ token Turkish corpus. Benchmarking the linguistic and multimodal capabilities of a model at this scale.

Mistral Small 3.224B paramsMultimodal20M tokens TRUnsloth

Hugging Face · Lodos-24B-Instruct-2510

Gradio Application2024–2025

Document extraction & classification via VLMs

Universal pipeline accepting any document type (CV, invoice, report) as PDF or image. Extraction driven by user-defined schemas.

GradioVLMsCustom SchemasPDF / Image

Applied VLM2026

Constat Amiable · VLM extraction

QLoRA on Qwen3-VL 8B to extract structured data from French insurance accident reports (constats amiables). Trained on synthetic data.

Qwen3-VL 8BQLoRAUnslothDocument AISynthetic data

Hugging Face · qwen3-vl-8b-constat-amiable-lora

R&D · Multi-agent system2026

Autonomous Company · an AI company that runs itself

A CEO agent hires workers, writes their prompts, assigns tools and budgets, then orchestrates the whole run. Workers talk peer-to-peer (A2A) and escalate to the CEO, who arbitrates with six decisions and reshapes the team mid-run: budget caps, a dangerous-command guard, secret redaction, 1500+ tests. Pointed at the Linux kernel, its autonomous bug-hunt found the defects behind two of my upstream-merged patches; I verified and wrote the reports.

CEO + workersA2A protocolClaude Agent SDKPython · uv1500+ tests

Technical competitions2024–2025

MCP & Agents · hackathon series

Building agents under tight time constraints, focus on interoperability.

MCPAgentsMistral AIHugging Face

Mistral AI × Cerebral Valley

In-person · Elite

Hugging Face Agents-MCP

Online

Hugging Face MCP 1st Birthday

Online

R&D project

Fine-tuning · 2025

Lodos-24B

Full fine-tuning of a 24-billion-parameter multimodal LLM on a 20M+ token Turkish corpus.

24BParameters

20MTR tokens

16kContext length

Base model: Mistral Small 3.2
Dataset: Turkish corpus · 20M tokens
Architecture: Multimodal · Vision + Text
Strategy: Full fine-tune · Unsloth + TRL
License: Apache 2.0

View on Hugging FaceHosted on Hugging Face · hasanbasbunar/Lodos-24B-Instruct-2510

Shipped product

macOS · 2026

Talkink

100% on-device push-to-talk dictation for macOS. Three ASR engines, zero cloud, shipped via Homebrew.

100%On-device

3ASR engines

0Cloud calls

Platform: macOS · Apple Silicon
ASR engines: Qwen3-ASR · Nemotron · Voxtral
Runtime: Apple MLX · 100% on-device
Distribution: Homebrew · Apple-notarised
Bonus: Native TTS · "Read Claude Code Aloud"

Visit talkink.app Source code

Open source · github.com/hasso5703/talkink

The lab

My own stack, in production.

Where I experiment: GPU homelab, self-hosted inference, public demos. What I build outside the day job.

Homelab

GPU homelab

DGX Spark GB10 128 GB and a gaming PC with an RTX 4070 Ti Super. LLM inference at home.

Asus Ascent GX10 · DGX Spark

Gaming PC · RTX 4070 Ti Super

Live demo

84 ♥

Voice Cloning · XTTS-v2

Few-second voice cloning. My most popular Hugging Face demo.

Live demo

13 ♥

SAM 3 · Segmentation

Interactive image segmentation (Segment Anything 3), as a Gradio Space.

Live demo

Nano Vision Studio

A studio for lightweight vision tasks, several models in one UI.

Open source

Briefing IA

Daily AI newsletter (3×/day): deterministic data → HTML pipeline, Claude Code routine + Resend.

Open source

NeuroDash

Real-time NVIDIA GPU telemetry in the browser. Single-file Python dashboard.

Skills / 04

The stack, from bare-metal to interface.

/ 01

Core Engineering

Python 3.10+
AsyncIO
Typing
Bash
CI/CD

/ 02

Systems · C

C
Linux kernel
modpost / kbuild
lore.kernel.org
Low-level debugging

/ 03

AI / Machine Learning

PyTorch
Transformers
LangChain
MCP
RAG
Agents

/ 04

Fine-tuning

Unsloth
TRL
LoRA / QLoRA
Full fine-tune
GRPO

/ 05

Inference

vLLM
llama.cpp
GGUF
MLX
Quantization

/ 06

Interfaces

Gradio
FastAPI
Swift / macOS
Next.js
REST APIs

/ 07

Infra · Homelab

DGX Spark
exo cluster
Docker
Cloudflare
Vercel

/ 08

Tools

Claude Code
Cursor
Git
Linux
Hugging Face

Education & Certifications / 05

Five years grounded in data & AI.

Education

Master OIVM
Optics, Image, Vision, Multimedia
Université Paris-Est Créteil
SDTS track · Distributed Systems & Data Science
2024 – 2026M2
BUT Data Science
Université de Lorraine, Metz
Honours · Top of class
2023 – 2024
DUT in Computer Science
Université de Lorraine, Metz
2020 – 2022
Scientific Baccalaureate
Lycée Georges de la Tour, Metz
2018

Continuous learning

Twenty-three certifications.

From basic ML to VLM post-training and the Model Context Protocol.

GRPO · Group Relative Policy Optimization

Hover to pause · drag to navigate

By the numbers

Receipts, not promises.

A few numbers from my last four years of building.

2patches merged into Linux upstream

32Bmodel parameters fine-tuned

30M+fine-tuning tokens processed

23professional certifications

3elite hackathons (Mistral, HF)

3spoken languages

Contact / 06

Let's talk about your next build.

Forward Deployed AI Engineer at Liberty International. Always up for talking agents, fine-tuning, or an ambitious project to build together.

basbunarhasan@gmail.com

I build autonomous agents.

Where models meet production.

The agent, alive.

Forward Deployed AI Engineer · Work-study

AI Engineer & Tech Lead · Work-study

Data Scientist · Internship

Machine Learning Developer · Internship

page_pool · memory-provider leak on error path

modpost · stack buffer overflow

scripts/sorttable · off-by-one

bpf_dbg · dual off-by-one

Lodos-24B · Full fine-tuning of a 24B multimodal LLM

Document extraction & classification via VLMs

Constat Amiable · VLM extraction

Autonomous Company · an AI company that runs itself

MCP & Agents · hackathon series

GPU homelab

Voice Cloning · XTTS-v2

SAM 3 · Segmentation

Nano Vision Studio

Briefing IA

NeuroDash

Core Engineering

Systems · C

AI / Machine Learning

Fine-tuning

Inference

Interfaces

Infra · Homelab

Tools

Master OIVM

BUT Data Science

DUT in Computer Science

Scientific Baccalaureate

Let's talk about your next build.