Pierre-Guillaume Sprimont

ASTROPHYSICIST · SENIOR SOFTWARE ENGINEER · SSA / SST EXPERT

Twenty years of engineering between the night sky and the machine room — from galaxy-cluster surveys and robotic telescopes to the European space-surveillance ecosystem, and now to fully on-device AI systems. I design software the way observatories are built: to keep working, unattended, for years.

> AI Projects

The heart of my current work. Two open platforms, both designed around the same conviction: your data, your model, your machine.

Engram — Local AI Platform for Android

Architect & Developer | on-device LLM · multi-APK platform

A complete AI platform running entirely on the phone: a LiteRT-LM engine (Gemma-family, 4-bit, GPU), a token-aware agent with tools, a long-term semantic memory with photos and EXIF metadata, personas, and a family of mini-apps (Vision, Voice, Summarizer, Memory). An optional OpenAI-compatible cloud mode bridges to DeepSeek and friends — opt-in only.

  • Engine + agent + tools + two HTTP servers, all in one foreground service.
  • Semantic / full-text / tag / temporal memory search; memories carry photos with date & GPS as separate metadata.
  • Preventive KV-cache management and silent overflow recovery — the chat never crashes.
  • Multi-APK architecture: every mini-app talks to the brain over a local HTTP API.

[ project page ] [ source ]

ExecAI — Sovereign, Memory-Driven AI Agent

Architect & Developer | Python · agentic framework v4

A modular agent framework that runs on your own local models (Ollama, llama.cpp) or in the cloud (DashScope/Qwen, DeepSeek) without changing a line of code — with a real long-term memory that learns from every session.

  • Hybrid RAG memory: ChromaDB + BM25 + temporal decay, automatic distillation of durable facts, deduplication/merging.
  • Hardened file tools (atomic writes, checksums, binary detection), context-overflow protection per backend, graduated safety modes (plan / edit / auto / PARANOID).
  • Full test-bench, MIT, 100% offline-capable on modest hardware.

[ project page ] [ source ]

Edge AI & Industrial Vision

C++ / Computer Vision Engineer
  • TensorFlow models for real-time people and object detection on Paris metro access-control gates (CAN Bus).
  • Industrial quality-control pipelines with Halcon and OpenCV, coupled with B&R motor control.
  • CUDA / TensorRT / quantization work for constrained hardware.

> Research in Astrophysics

Université de Liège — PhD Fellow & Scientific Attaché (2003–2011)

Belgium · Prof. Jean Surdej’s team

Observational cosmology and software for the XMM-LSS survey — one of the deepest X-ray surveys of the large-scale structure of the Universe — and the ILMT (International Liquid Mirror Telescope) project. I built the C++ spectroscopy and imaging pipelines that turned raw telescope frames into published science: optical follow-up of X-ray-selected galaxy clusters and AGN, faint-object photometry, web interfaces for data distribution.

  • DEA (postgraduate diploma) in Astrophysics — C++ spectroscopy software, 2005.
  • Co-author on three Astronomy & Astrophysics papers (XMM-LSS clusters, AGN classification, the quadruply-imaged quasar HE 0435–1223).
  • La géométrie de Lobachevski — Revue des Questions Scientifiques, 2011.

INAF / IASF Bologna — Research Fellow (2012–2015)

Italy · GLORIA project

Robotic-telescope interfaces, astronomical camera drivers and an astrophysical image database, with P2P data sharing and HTML5/WebGL visualisation — astronomy for the web, a decade before everyone else noticed. In parallel: GRBspec at IAA-CSIC (Granada), the gamma-ray-burst spectroscopic follow-up database, and the NILDE inter-library system at CNR Bologna.

Outreach

Belgium · Italy · Romania

Public sky-watching events and astronomy outreach since 2002 — including teaching Earth & Life Sciences at the Lycée Français Anna de Noailles in Bucharest (2016–2017).

> Space Surveillance & Tracking

GENDARED — Generic Data Reduction Framework

Lead Architect & Principal Developer | GMV, 2020–2026

Principal developer of the data-reduction, detection and tracking framework for orbital objects observed from optical telescopes. Prototyped under the ESA Romanian Industry Incentive Scheme, GENDARED became a reference pipeline of the European space-surveillance ecosystem — cited in international peer-reviewed publications.

  • WRLagez1 (EU SST) — delivery to OHB Digital and the German military SST Centre.
  • GRNOC, RONOCOPS, ROSSA — software delivery and data-processing services within the COSST consortium.
  • DEVISE — ESA study on faint-object analysis.
  • Presented at the 8th European Conference on Space Debris (ESA/ESOC, 2021).

> Key Skills

Languages & Frameworks

C++ (Expert) CUDA Python Java JavaScript / Node.js Fortran Android (Kotlin)

AI & Computer Vision

TensorFlow PyTorch LLM Integration Agentic Systems LiteRT / on-device inference OpenCV Halcon YOLO TensorRT

Space & Astrophysics

GENDARED SSA / STM Astrometry Photometry Telescope Interfaces Data Reduction Pipelines

Systems & Hardware

GNU/Linux Admin TCP/IP · WebSocket · WebRTC Microcontrollers CAN Bus SQL · MongoDB

> Languages

French (native) · English (C1/C2) · Romanian (C1/C2) · Italian (B1/B2) · Dutch (B1/B2)

> Curriculum Vitae

> Contact