Projects

A pure-Python simulation of a 1947-era germanium BJT transistor: Ebers-Moll large-signal model, switch and amplifier analysis, and a physical comparison with the vacuum tube it replaced.

June 12, 2026

A faithful simulation of ENIAC's decade-counter arithmetic and the ballistic trajectory calculations the machine was actually built to compute.

June 6, 2026

Python implementation of Vannevar Bush's Memex concept from 'As We May Think' (1945). Demonstrates associative trail navigation, document linking, and the contrast between index-based keyword search and curated associative browsing: the logical precursor to hypertext and the web.

June 1, 2026

Python simulator of the stored-program architecture described in von Neumann's "First Draft of a Report on the EDVAC" (1945). A 256-word memory machine with nine opcodes demonstrates the fetch-decode-execute cycle, conditional branching, and the bottleneck that still haunts modern CPUs.

May 26, 2026

A Python demonstration of the combinatorial explosion argument from the 1973 Lighthill Report, using the 15-puzzle as the benchmark domain. Implements breadth-first search (exponential growth), A* with Manhattan distance heuristic (Hart, Nilsson, Raphael 1968), and iterative-deepening A* (Korf 1985). Produces a side-by-side comparison table of nodes expanded at increasing solution depths, making Lighthill's core mathematical argument visible and measurable.

May 24, 2026

A Python implementation of MYCIN's core inference engine: production rules with certainty factors in [-1, 1], backward chaining over a goal tree, and the Shortliffe-Buchanan certainty combination formula. Includes a simplified bacteraemia knowledge base derived from the 1979 MYCIN evaluation domain, demonstrating ranked organism diagnosis and antibiotic recommendations across three clinical case scenarios.

May 24, 2026

Python implementation of the Lorenz SZ-40/42 cipher machine used by Hitler's high command, including the statistical delta attack that Colossus performed to find chi wheel settings. Demonstrates the "depths" XOR key-cancellation trick from the Athens/Vienna 1941 mistake.

May 21, 2026

Python implementation of the 1943 McCulloch-Pitts formal neuron model. Builds AND, OR, NOT, NAND, and XOR gates from binary threshold neurons, implements a temporal sequence detector (pattern 101), and contrasts with a Rosenblatt perceptron.

May 18, 2026

Full Python simulation of the Wehrmacht 3-rotor Enigma machine and Turing's Bombe attack. Demonstrates Enigma's fatal self-map flaw, crib dragging, and recovery of rotor start positions from known plaintext.

May 15, 2026

Python simulation of Konrad Zuse's 1938 Z1: the world's first programmable binary floating-point computer. Implements Z1's 22-bit float format, gate-level ripple-carry adder, bistable memory cells, and program tape executor.

May 11, 2026

A Python implementation of Shannon's 1937 algebraic model of switching circuits. Composable gate primitives, truth table generation, SOP/POS canonical forms, Shannon cofactor expansion, and a ripple-carry adder, all from first principles.

May 7, 2026

A Python implementation of dense vs. sparse attention patterns — sliding window, LSH-bucketed, and top-k learned routing — with a retrieval recall benchmark that shows why prior sparse attention approaches degraded at scale and what a learnable router buys you.

May 5, 2026

A faithful Python implementation of the formal Turing machine model from the 1936 paper. Includes parity checker, palindrome recogniser, unary adder, and a demonstration of the Universal Turing Machine concept via machine encoding.

May 2, 2026

A two-layer neural network trained with hand-rolled backprop in NumPy — forward pass, loss, analytic gradients via the chain rule, and a parameter update loop — no autograd, no magic.

April 25, 2026

A self-contained transformer encoder built from scratch in PyTorch: multi-head scaled dot-product attention, positional encoding, and a feed-forward sublayer — under 150 lines with annotated shapes at every step.

April 18, 2026

Vanilla SGD, momentum, RMSProp, and Adam implemented from scratch in NumPy — with a shared loss surface visualisation so you can watch each algorithm find the minimum differently.

April 10, 2026

End-to-end fine-tuning of bert-base-uncased on SST-2 using the HuggingFace Trainer API. Covers tokenisation, TrainingArguments, evaluation with the GLUE metric, and a ready-to-use inference pipeline.

January 13, 2020

Loads bert-base-uncased and extracts per-layer, per-head attention weights for any sentence. Includes a matplotlib heatmap showing how token pairs attend to each other across all 12 layers.

January 6, 2020

Spring XML bean wiring, constructor injection, and a JUnit test that needs no application server — showing exactly why Spring made Java testing practical.

June 22, 2003

A Hibernate 2.x mapping file, POJO entity class, and Session usage for loading, saving, and querying — the complete working example from the article.

July 22, 2001