compusential
About & Perspectives

Yang Pei

📍 Based in London, UK

I grew up in Mannheim, Germany for my first ten years before moving to Beijing, where I attended the Western Academy of Beijing (WAB) and Beijing BISS International School. Starting school with only basic classroom English was a bit of a shock, but with some help from a Dutch classmate who spoke German and friends from the school bus, I quickly found my footing. That period in Beijing was also where my obsession with hardware took off—building PCs around AMD Athlon CPUs (and occasionally burning through them), flashing firmware, and playing games.

At eighteen, I moved to the UK on my own to study Computer Science at the University of Reading. Those early months in a cold house share with two other students were solitary, but the years of hands-on hardware tinkering paid off eventually. When I interviewed for a junior web dev role at NVIDIA's European office in Theale, being able to rattle off pretty much any GPU spec got my foot in the door.

Today, I work as a Technical Development Manager at LGC Group supporting NIHR, building production AI platforms and agentic workflows, while running a local compute lab with an NVIDIA DGX Spark (128GB unified VRAM) and an RTX 4090 to experiment with unconstrained open-weight models.

Outside of software, I've spent years shooting with Leica rangefinders—both film and an M10 digital body, usually paired with a 35mm Summicron ASPH or 50mm Summilux ASPH. To me, photography isn't about chasing megapixel count; it's about capturing the essence of a subject with honest glass and natural light.

I built Compusential as a personal corner on the web for things I find interesting: AI and computing architectures, photo essays, and practical tech explorations.

Engineering Log

Systems & Recent Builds

Select a build to expand architecture & field notes

Local Silicon & SynthesisSep – Oct 2026

compu-master

Self-Hosted Research Companion on Local Silicon

A self-hosted research companion running on a 128GB NVIDIA DGX Spark. Searches a local YaCy index, extracts verbatim quotes from source pages, and uses local vLLM models to compile cited study notes and PDF briefings with zero API credits.

DGX Spark (128GB)vLLMYaCySQLiteNext.js
Architecture & engineering notes▾

Architecture

  • Queries a self-hosted YaCy search node, strips page clutter, and saves verbatim quotes before calling the local model.
  • Drafts structured study notes via local vLLM and verifies every statement back against the saved source quotes.

Key Engineering Friction & Fix

Local reasoning models often spent their entire token budget inside thinking blocks before closing structured output; turning off extended reasoning for straightforward extraction calls kept the pipeline reliable.

Cloud Run & Spatial SystemsOct 2026

CityBreaker

Historic Pubs, Landmarks & 22-City Break Planner

An interactive 22-city travel map and weekend break planner with 10,800+ verified pubs, independent restaurants, thermal spas, landmarks, and nearby hotel stays, running on Google Cloud Run with a Cloud Storage-backed SQLite database.

Next.jsMapLibre GLCloud Run + SQLiteGemini Flash-LiteOpenStreetMap
Architecture & engineering notes▾

Architecture

  • In-container SQLite database backed by a Cloud Storage volume to avoid standing monthly database bills.
  • Twice-daily background curation job using OpenStreetMap and travel creator playlists with Gemini Flash-Lite to add new venues and remove closed ones.

Key Engineering Friction & Fix

Existing hotel booking cookies in a visitor’s browser occasionally overrode text-only destination redirects; passing explicit latitude and longitude coordinates alongside the venue name fixed it.

Interactive EducationOct 2026

Stella Academy

Interactive Learning Set Against the Story of Space

31 free interactive lessons for ages 9 to 18 (key stages 2 and 3, GCSE, and A-level), matched to UK Department for Education curriculum statements and arranged as five space missions.

Next.jsTypeScriptInteractive SVG & CanvasNASA Media APIPlaywright
Architecture & engineering notes▾

Architecture

  • Five space missions (/journey) with browser-saved progress, NASA mission footage, and a browser workbench (/lab).
  • Pairs each interactive model with a step-by-step worked example and a three-round practice quiz.

Key Engineering Friction & Fix

Started at a hackathon with an AI chat tutor and login wall, which got in the way of learning. Retiring the chatbot and auditing the AI-drafted notes line by line caught US curriculum defaults, formatting bugs, and subtle explanation errors.

Empirical ML & Local ComputeSep – Oct 2026

Synthetic Biomedical Abstracts (Replication Study)

18-Replicate Evaluation of Local Qwen3.6-35B vs Gemini 2.5 Flash

An empirical study testing whether synthetic abstracts generated on local silicon (Qwen3.6-35B) vs Gemini 2.5 Flash can train a classifier that generalises to real held-out NIHR project summaries.

Qwen3.6-35B (Local vLLM)Gemini 2.5 FlashBioMed-RoBERTaPyTorch
Architecture & engineering notes▾

Architecture

  • Compared pure synthetic substitution against synthetic augmentation across 18 independent training runs.

Key Engineering Friction & Fix

Variance across independent runs proved three times wider than in-process random seeds, showing why 3-seed ML benchmarks often overstate results. Across 18 runs, augmenting small real datasets with local Qwen3.6-35B matched Gemini 2.5 Flash.

Agentic Safety & RAG2026

Guardrails AI Hub Benchmark & Omni-Evidence

Hybrid Output Moderation & Multi-Agent Evidence Synthesis

A 65-validator Guardrails AI evaluation testbed running against local open-weight models, alongside a Google ADK multi-agent evidence synthesis engine with offline fallback.

Google ADKGuardrails AIMicrosoft PresidioVertex AI GeminivLLM
Architecture & engineering notes▾

Architecture

  • Combines fast CPU scanners (PII and secret detection) with selective model-based checks for ambiguous outputs.

Key Engineering Friction & Fix

Running a model judge on every response adds too much latency; handling obvious checks on CPU first keeps response times practical.

Enterprise Production AI2025 – 2026

NIHR Enterprise AI Platforms

Production AI Portal, Evidence Retrieval & Offline GPU Extraction

Production generative AI tools built and shipped end-to-end: a multi-model caseworker portal, a RAG evidence retrieval platform, and an offline GPU extraction pipeline that saved ~400 hours of manual grant review.

GCPVertex AI & ClaudeVector SearchLocal GPU PipelineRBAC
Architecture & engineering notes▾

Architecture

  • Multi-provider chat, vector retrieval, role-based access control, and batch GPU extraction for long-form funding applications.

Key Engineering Friction & Fix

Shipping to real caseworkers showed that predictable output schemas, clear permissions, and verifiable citations matter far more than autonomous agent loops.

Visual Frames & Photography

Moments captured through the lens across Germany, Canada, the US, and the UK.

Explore the gallery →

Get in touch

Always open to conversations around computing architectures, photography, or interesting engineering challenges.