Single-Tower High-Performance Rig

This machine currently runs 70B+ parameter models locally and processes terabyte-scale microscopy datasets. The full build-out to three GPUs and 1TB of RAM will generate value every day and stay relevant as I upgrade components.

The core philosophy here is modularity without compromise. The Threadripper 7970X gives me 48 PCIe 5.0 lanes, meaning I can run three flagship GPUs at full x16 bandwidth without needing £GBP thousands more on a dual-socket EPYC or enterprise Xeon platform. The ASUS Pro WS TRX50-SAGE has IPMI and ECC support, so this isn't just a gaming rig with delusions of grandeur—it's production-grade infrastructure.

This started as a microscopy research platform. I needed something that could crunch through days of high-resolution imaging data while simultaneously running LLM inference for automated experimental workflows. Turns out, the Venn diagram of "can process microscopy data" and "can run massive language models" is just "absurdly powerful computer."

ECC memory isn't negotiable. When you're running multi-day experiments or training on scientific data, a single bit flip can corrupt everything. The 128GB DDR5 ECC setup expands to 1TB because why set artificial limits?

The 96TB ZFS pool means I can stop worrying about cloud storage costs and actually work with real datasets locally. Checksumming, snapshots, data integrity—all the things you want when your data represents months of work.

The RTX 6000 Blackwell (96GB VRAM) is cutting-edge enough to stay relevant as models get bigger, and when I need more, I'll just add two more GPUs. 288GB total VRAM in a single tower. No racks, no datacenter, no monthly AWS bills making me cry.

Scientific PDF Extraction Pipeline

A production pipeline for extracting high-fidelity text from scientific books and papers — without OCR. Traditional OCR collapses on the content that matters most: multi-column layouts, inline LaTeX equations, figure captions interleaved with text, and dense technical notation. This pipeline bypasses that entirely by treating each page as an image and using a vision-language model to read it directly.

microscopy_indexing_yolo-qwen2-72b ↗

DocLayout-YOLO (YOLOv10, trained on DocLayNet) runs first and classifies every region on the page into six content classes: Text, Title, Section-header, List-item, Caption, and Formula. Each detected region is cropped and passed individually to Qwen2-VL-72B, which transcribes the image directly. Tall blocks are chunked vertically before inference and reassembled afterward. Multi-column reading order is reconstructed from bounding box X-coordinates — left column top-to-bottom, then right column.

The extraction loop writes a checkpoint after every page. If the process is interrupted — OOM, power loss, deliberate pause — re-running the script picks up exactly where it left off. Batch size is reduced automatically on OOM and cautiously ramped back up on recovery. VRAM is monitored via pynvml against a configurable threshold; cleanup runs at page boundaries, not per-block, which is a meaningful throughput win on 72B models.

Delphi Deploy

A complete, self-contained deployment stack for Qwen3.5-122B-A22B (MoE, 22B active parameters) on a single high-VRAM GPU. Built around vLLM with an OpenAI-compatible API, it bundles a CLI coding agent, an interactive chat interface, a Rust-based web GUI with session management, and browser automation — all wired to the local model with zero cloud dependency.

The stack is named after the Oracle at Delphi. The model persona, Delphi, greets every new conversation as the Pythia at the Temple of Loxias — a private oracle running entirely on local hardware.

One-command bootstrap. bootstrap.sh is fully idempotent — it installs the conda environment, downloads ~79GB of model weights from HuggingFace, builds the Rust binary, caps GPU power via systemd, and starts the server. Re-running it at any time skips completed phases. A single config file, delphi.conf, is the source of truth for every runtime parameter; nothing is hardcoded elsewhere, and a drift scanner enforces this.

Claw is a modified Rust reimplementation of the Claude Code CLI, adapted to run against the local vLLM endpoint. Modifications add session load/resume from the web GUI, inline turn stats (elapsed time, tok/s), and an external HTML GUI file served at runtime — editable without recompiling.

The chat interface streams responses with live thinking-mode word counts, a context bar that tracks token usage against the 128K window, and automatic context compaction — when older messages push token usage past a threshold, the model summarises them into a dense block, preserving context indefinitely without truncation.

Epiproc Atlas — Procurement Intelligence Platform

A multi-tenant SaaS platform that turns scientific-supplier invoices into procurement intelligence. It takes the vision-LLM extraction approach from my Scientific PDF Extraction Pipeline — pages read as images by a vision model, never OCR — and wraps it in a full production stack that is sector-agnostic and scalable: schema-per-tenant PostgreSQL, server-side sessions, MFA, audit logging, a crash-safe job queue, observability, and a Cyber-Essentials-ready compliance pack. It runs entirely on the single-tower rig — UK data residency, no cloud.

Every database query is scoped to the tenant schema via SET search_path, and a CI test creates two tenants and asserts that one can never read the other's rows. Reports are generated on demand by either a self-hosted local model (GPU) or a cloud API, dispatched through a Postgres-backed job queue and worker that survives server restarts.

Screenshots show a fully anonymised demonstration tenant — supplier names, departments, products and figures are fictitious.

Quercus europaea

My personal theory about the rise of agriculture is that oaks are the cornerstone species that enabled the revolution, not cereals. Oak species are ubiquitous in every region where agriculture arose independently: the Fertile Crescent in the Levant; Mesoamerica, home to the highest diversity of oak species on Earth; and the Yangtze and Yellow river basins in China. Acorns are abundant in the archaeological record, and only in the Fertile Crescent does the balance tilt slightly towards cereals as the dominant crop. Both China and Mesoamerica tilt it towards acorns as precursors, before corn, rice and millet reached domestication. Japan and California had acorn economies of such scale that they never needed to develop agriculture at all.

Why? Oaks are deep-rooted trees, less sensitive to seasonal variation, and they respond to those fluctuations by masting: every four or five years a bumper crop is produced. Acorns are storable. The literature discusses how storage pits and the stone tools believed to be for cereal processing could have originated as acorn processing, or served both products seasonally. A preservation bias towards cereal grains may have handed us the wrong archaeological conclusion.

It took me eight months to collect the twenty specimens on this sheet.

The project started in Bulgaria. The country has amazing forests and beautiful mountains, and with the Giro d'Italia kicking off there I had the perfect excuse for a trip to visit my dear friend Georgi Merev. While the cyclists were having a tough start in rain and cold, I was roaming the streets, parks and woodlands around Veliko Tarnovo and Plovdiv, observing and collecting.

Early summer took me to Sardinia, a Quercus lover's paradise. The maquis and garrigue are incredible landscapes, and it is there that the altitudinal gradient shows oaks specialising. The lowlands are the realm of Q. ilex, the holm oak, while the upper slopes are dominated by Q. pubescens sensu lato. Yes, Q. pubescens is a tricky species: the variation is so large that wherever it occurs botanists have created all sorts of subspecies, or simply described new species. Introgression and hybridisation are rampant, so sensu lato it is. But the interesting part is in the highlands, away from the coast. There the sughera is king (Q. suber), and the composition of these cork woodlands, which double as animal pens, is outstanding.

From Sardinia I travelled to the Netherlands, an awkward place to look for oaks, but if anything defines the modern Dutch landscape it is that this is the land of arborists and florists. The streets of any Dutch city are small arboretums, and simply walking through them reveals a majestic mixture of American red oaks and European white oaks. The Hortus Botanicus in Amsterdam has two jewels: a fine cork oak planted next to the shop, and a beautiful Q. × turneri. While I collected leaves the Breda Jazz Festival was in full swing, so the botany came enriched by classical and contemporary jazz, dotted with a good dose of salsa.

Finally, the UK is a huge arboretum disguised as a country. Because gardens and wealth have lived together for so long on these islands, you get the best of both worlds: classical cottage gardens, and the monumental private collections of rare specimens in the stately homes. A visit to Wimpole Estate hands you Q. ilex, Q. cerris and Q. petraea (rare in East Anglia) on a silver plate. A visit to the University of Surrey campus rewards you with a superb Q. pyrenaica, one of perhaps three in the whole country. Walking the streets of Cambridge you find all sorts of treasures hidden in plain sight — a beautiful Q. cerris on the rugby grounds, two fine Q. ilex in the gardens of Greenwich House, not to mention the trove in the Botanic Garden.

Eighteen photographs from the collecting trips. Click any image for the full-size version.

The plate is ordered by section, not by size or appearance. A section is the rank between genus and species: a lineage of oaks descended from a common ancestor and sharing a set of characters — leaf type, tooth type, bark, and how long the acorn takes to ripen. It is a more useful ordering than appearance because it predicts behaviour. Oaks cross freely with other members of their own section and only rarely outside it, so section boundaries do the work that species boundaries fail to do in this genus. The three sections here belong to two subgenera: sect. Quercus in subgenus Quercus, sects. Cerris and Ilex in subgenus Cerris.

The two Q. ilex specimens are included because juvenile foliage is spinose and adult foliage entire, and the two are routinely logged as separate taxa in field records. Q. × rosacea (robur × petraea) is common in Britain and frequently recorded as one or other parent.

All specimens are reproduced at true mounted size against a single 100 mm bar. No specimen is enlarged or reduced to fit its position, so relative size across the plate is measurable: Q. frainetto at the top left and Q. coccifera in sect. Ilex differ by roughly an order of magnitude in lamina area.

One map per species, twenty in total, at the foot of the plate. Equal-area projection, native range only, no political borders. Fill colour carries the section: olive for Quercus, rust for Cerris, teal for Ilex. Q. ilex and Q. pubescens are mapped sensu lato, as both are aggregates. Ranges are approximate and intended for orientation.

Sectional treatment follows the current placement of Quercus in subgenera Quercus and Cerris. Aggregates are marked s.l. Where a subspecies was collected it is given in full (Q. infectoria subsp. veneris). Hybrid parentage is stated as collected or as determined from morphology, not from genotyping.