Skip to content
View Marc-Dvci's full-sized avatar
9 followers · 0 following

Block or report Marc-Dvci

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Marc-Dvci/README.md

Marc Donovici

AI implementation lead for Group Internal Audit at Crédit Mutuel Alliance Fédérale in Strasbourg. I take generative AI from a business problem to production: six solutions shipped to a 500-person function, with 80%+ regular usage. Outside work I build offline, multimodal and agentic systems in AI competitions.

LinkedIn

Results

Result Competition Project
3rd of 850+ teams Google MedGemma Impact Challenge 2026 FieldScreen AI · writeup · Google blog
Winner, EdgeAgent track Global AI Hackathon with Qwen Cloud (Alibaba Cloud) LUMEN · demo
Finalist, 1 of 6 European Patent Office CodeFest 2026 IP Value Framework
2nd place, Pro category Euro-Information Concours de Code 2025 (Python, C++)
Honorable mention Tesseract 2026 Hackathon Vitrify

Selected projects

Offline healthcare AI

  • FieldScreen AI: tuberculosis screening for community health workers with no network. Fine-tuned MedGemma on chest X-rays, cough audio classification, voice intake and output in 15 languages, on one laptop.
  • LUMEN: neonatal danger-sign screening on the WHO IMNCI protocol. Runs on Qwen Cloud and falls back to a fine-tuned on-device model when the connection drops. Android app.

Governed agents for regulated work

  • AssuranceOS: an internal-audit platform run by 19 governed agents on Gemini. Signed releases, a default-deny tool gateway and evidence validation.
  • Lore: a permission and lineage control plane for archival media on ClickHouse. Every query family answers in under 33 ms at p99 on 300M rows.
  • Threshold: several care organisations answer one person's AI assistant through WebMCP and compose a single plan. Video

Performance and systems

  • fastpath64: an Arm smmla kernel for IQ4_XS in llama.cpp. Prefill goes from 0.58x to 1.24x of Q4_K on Neoverse N2.
  • decimal-rs: decimal.js ported to Rust. The original test suite runs unmodified against the Rust build.
  • Vitrify: inverse design of organ nanowarming as one differentiable function across JAX and FEniCSx.

Stack

Python · TypeScript · Java · Rust · C · FastAPI · React · Google ADK · AWS Strands · MCP · Vertex AI · Cloud Run · AWS CDK · Alibaba Function Compute · ClickHouse · Docker · pytest · Playwright

Pinned Loading

  1. FieldScreen_AI FieldScreen_AI Public

    Offline TB screening for community health workers: MedGemma chest X-ray, cough audio, voice intake, 15 languages. 3rd of 850+, Google MedGemma Impact Challenge 2026.

    Python 4 6

  2. Vitrify Vitrify Public

    End-to-end inverse design of organ nanowarming across JAX and FEniCSx.

    Python 1

  3. All-Access All-Access Public

    Recover a disrupted film shoot without ever dropping an approved access arrangement — and prove the plan reached everyone. Agentic Cinema · IBM track.

    Python

  4. AssuranceOS AssuranceOS Public

    Adaptive, AI-Native Internal Audit Platform

    Python

  5. fastpath64 fastpath64 Public

    2.12x faster IQ4_XS prefill on Arm Neoverse: the missing smmla repack kernel for llama.cpp. IQ4_XS is the smallest standard 4-bit GGUF and had an accelerated matmul path on Intel AMX but none on Ar…

    Python

  6. LUMEN LUMEN Public

    Neonatal danger-sign screening for community health workers. Qwen Cloud with an on-device fallback. Winner, EdgeAgent track, Qwen Global AI Hackathon.

    Python 2