Skip to content
View gabrielkpo's full-sized avatar

Highlights

  • Pro

Block or report gabrielkpo

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. 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
gabrielkpo/README.md

Hi 👋🏼

AI engineering student at ENSEEIHT · ML engineer at Finary

🔨 Now

Four projects on Mistral models, built in public:

  • SecoursTech: RAG assistant over official French firefighter doctrine. Every answer cites the guide and page; it refuses when the doctrine is silent. Ingestion and Postgres index done, retrieval and generation next.
  • Lifter: training coach agent on my own Garmin data (LangGraph + Mistral), with human approval before any write. v0.6.0 in production.
  • Alias: local PII masking before an LLM call (FR + EN). Rules, masking and restoring done, NER model trained (mmBERT-small); combining rules and model next.
  • hpc-inference-sizing: how many GPUs to serve Mistral Small 4 (119B MoE) under a latency target, measured with vLLM on a rented multi-GPU node. Protocol written, memory model and measurement bench next.

🚀 Projects

Project What it is Stack
garmin-mcp-remote Self-hosted MCP server that connects Garmin Connect to claude.ai, behind an OAuth 2.1 + PKCE layer, hardened and tested Python · Docker · nginx · OAuth
Fall-detection Fall detection from body pose (MediaPipe + 1D CNN), honestly evaluated (frozen test split, bootstrap intervals), served by an API and an in-browser demo, with drift monitoring Python · PyTorch · FastAPI · ONNX
video-super-resolution-fsrcnn School research project: classical upscaling vs sparse coding vs FSRCNN, plus a temporal filter that skips unchanged blocks to save compute Python · PyTorch · OpenCV
augmented-rag Research fork of gbrain: retrieval measured on a real 1,200-page knowledge base, each change kept or dropped on benchmark evidence TypeScript · Python · Postgres

🧰 Stack

Python · PyTorch · FastAPI · LangGraph · vLLM · PostgreSQL · Docker · GitHub Actions · Git · Linux

📫 Contact

LinkedIn · Portfolio

Pinned Loading

  1. garmin-mcp-remote garmin-mcp-remote Public

    Turnkey remote-hostable Garmin Connect MCP server (Taxuspt/garmin_mcp) with bearer-token auth in front

    Python 2

  2. alias alias Public

    Local PII masking before an LLM call: detect personal data (FR + EN), replace it with stable aliases, restore it in the answer.

    Python 2

  3. lifter lifter Public

    Training coach agent on Garmin data (LangGraph + Mistral), with human approval before any write

    Python 2

  4. secours-tech secours-tech Public

    RAG assistant over official French firefighter doctrine (GDO, GTO, SSUAP, PSE): answers with cited document and page, refuses when the doctrine is silent

    Python 1

  5. augmented-rag augmented-rag Public

    Forked from garrytan/gbrain

    Research fork of garrytan/gbrain: retrieval measured on a real 1,200-page knowledge base, with reranker fusion, doc2query and supersession, each kept or dropped on benchmark evidence.

    TypeScript 1

  6. hpc-inference-sizing hpc-inference-sizing Public

    Measured GPU sizing for LLM inference: how many GPUs a latency target needs and how many users a fixed fleet serves, from vLLM load tests of Mistral Small 4 on 2 and 4 NVIDIA A40s

    Python 2