~/projects
projects/
- drwxr-xr-xCarrier Audit: Freight Invoice Reconciliation[Python, FastAPI, React, TypeScript, Claude API, Playwright]2026-10
Checks a freight carrier's invoice against the contracted rate card, re-prices every shipment, and sends anything needing judgement to a person with the evidence attached. Code does the pricing and the decisions; the model only names each difference. 93.8% accuracy with zero false disputes on a 65-line synthetic evaluation set. Not deployed.
- drwxr-xr-xLearner Inbox Triage: Bilingual Message Routing[Python, Claude API, OpenAI-compatible APIs, Ollama, Pydantic, pytest]2026-10
Sorts learner email for an English and French training program into the right queue. A language model labels each message, and my code validates the label and picks the queue, so the model never replies on its own. On a local 3B model it routed 97.5% of a 40-message synthetic evaluation set correctly. Not deployed.
- drwxr-xr-xMnemos: Redis-Compatible Server[C++23, CMake, epoll/kqueue, MCP, Kafka protocol]2026-09
An in-memory key-value server written from scratch in C++23 that speaks the Redis protocol, so any Redis client works with it unchanged. Differential tests compare every reply byte for byte against real Redis. It also ships an MCP server that lets an AI tool inspect a live cache, and a Kafka broker that real Kafka clients can use.
- -rw-r--r--Evidence and Verification Server for an LLM Document Workflow[Python, MCP, JSON-RPC 2.0, Standard Library]2026-09
An MCP server, built with only the Python standard library, that lets a language model look up what evidence exists and what a claim may say across 95 documents. A deterministic checker then enforces the rules on the final PDF, so the model cannot skip one. It found all ten false claims a manual audit had found by hand.
- -rw-r--r--Hermes Agent Lab: Open-Source Agent Runtime on My MCP Server[Hermes Agent, MCP, Ollama, Local LLM]2026-10
I set up Hermes Agent, an open-source AI agent runtime, on a local model and connected my own MCP server to it in read-only mode. I checked every result independently: the server blocked every write, yet in one run the agent reported a change that never happened. I wrote up five failures with logs.
- -rw-r--r--Phishing Email Detection: Baseline, Zero-Shot and Fine-Tuned Transformer[Python, scikit-learn, PyTorch, Hugging Face, ModernBERT]2026-10
Compared three phishing email detectors on one fixed split of a public dataset of 17,513 emails: a TF-IDF baseline, and a ModernBERT-based model used zero-shot and fine-tuned. On the test split the fine-tuned model missed 7 of 981 phishing emails against 19 for the baseline, at a similar false-alarm count.
- -rw-r--r--Minutes: Team Meeting-Notes GraphQL API[Python, Django, GraphQL, PostgreSQL, pytest]2026-10
A Django and GraphQL API for team meetings, agendas, notes and action items, with per-object permissions so a non-member cannot tell whether a meeting exists. I cut the database queries for a nested read from 701 to 9 at 100 meetings, guarded by a regression test. Tested in CI against PostgreSQL 16.
- -rw-r--r--Network Outage Simulator[Java, Spring Boot, REST Assured, C#, .NET 8, Avalonia]2026-11
A Spring Boot REST API that simulates how a network breaks apart as nodes fail, reporting the size of the largest connected piece after each removal. Seeded runs are exactly reproducible, so attack strategies can be compared fairly. A C# desktop client drives it, and 11 REST Assured tests check the running server.
- -rw-r--r--Code Judge with a Lease-Based Worker Queue[Java, Spring Boot, C#, .NET 8, React, Vite]2026-10
A small online code judge. A Spring Boot control plane hands submissions to separate worker processes through expiring leases, so work returns to the queue if a worker disappears. A C# worker grades each submission and reports a verdict, and a React UI shows status live. It never runs untrusted code.
- -rw-r--r--Signature-Verified Webhook Receiver on Azure Functions[Python, Azure Functions, Azure Key Vault, GitHub Actions, Terraform]2026-09
A Python webhook receiver I deployed to Azure Functions that rejected any request with a bad HMAC signature, verified on the live endpoint. The secret was read from Key Vault through a managed identity, and GitHub Actions deployed through OIDC with no stored credentials. I later rewrote the stack as Terraform, applied it cleanly, then tore it down.
- -rw-r--r--Netdismantle: Open-Source Python Library[Python, NetworkX, pytest, mypy, Open Source]2026-08
The network-dismantling algorithms from my M.Sc. thesis, packaged as an installable Python library with a NetworkX-native API, a command-line tool and a documentation site. 317 tests at 100% line and branch coverage, with CI across Python 3.10 to 3.13 on Linux, macOS and Windows.
- -rw-r--r--M.Sc. Thesis: Network Dismantling in Large-Scale Complex Networks[Graph Algorithms, Network Science, Python, NetworkX, M.Sc. Research]2026-08
Finds a small set of nodes whose removal breaks a large network into small disconnected pieces. My two-phase algorithm breaks every cycle first, choosing nodes by Collective Influence, then splits what remains. It is competitive on sparse graphs and loses its edge as graphs get denser. Defended 19 August 2026.
- -rw-r--r--Open-Source Contribution: GitLab Orbit Knowledge Graph[Rust, TypeScript, tree-sitter, Open Source]2026-09
Fixed a bug in GitLab's Rust code-graph indexer, merged into main on 2 September 2026. "Who calls this?" queries came back empty for TypeScript because class fields were being dropped; my fix records them. Proved with a fails-before, passes-after check and six integration cases, then revised through maintainer review.
- -rw-r--r--Multi-Source Job Data Aggregation Pipeline[Python, Standard Library, TF-IDF, curses]2026-09
A dependency-free Python pipeline that collects job postings from seven applicant-tracking systems, including Greenhouse, Lever and Workday, into one format. It rate-limits each host, survives a failed source, and logs why every posting was filtered. Survivors are ranked by TF-IDF similarity to a resume.
- -rw-r--r--Colour-Aware Composed Image Retrieval[Python, PyTorch, CLIP, Mistral 7B, LoRA, HNSW]2025-08
Finds a fashion item from a reference image plus a text change, on the 37K-image FashionIQ dataset. A colour pre-filter removes about 97% of candidates before fine-tuned CLIP embeddings rank the rest, about four times faster. A Mistral 7B model fine-tuned with LoRA turns a free-text request into the search query.
- -rw-r--r--Terminal Text Editor (In Progress)[C, Terminal I/O, In Progress]2026-05
Working through a terminal text editor in C from scratch, with no external libraries, following antirez's kilo: raw-mode terminal input, VT100 escape codes for drawing the screen, a row-based text buffer, incremental search and syntax highlighting.
- -rw-r--r--Personal Portfolio Website[Next.js, TypeScript, Tailwind CSS, Resend, Vercel]2026-05
This site, built with Next.js 16, TypeScript and Tailwind CSS as a terminal you can type into (try help or cd projects). Projects and blog posts are files in the repository, added from the command line, and the contact form sends mail through Resend. Deployed on Vercel.
- -rw-r--r--Deep Graph Clustering of Spatial Transcriptomics (Mouse Brain)[Python, PyTorch, Graph Neural Networks, Scanpy, NetworkX]2025-04
Divided a mouse brain tissue section into anatomical regions from gene expression alone, with no labels, by clustering a graph of 2,800 tissue spots. In a three-person team I led the GCN and GAT models. Attention-based models followed the real anatomical boundaries most closely.
- -rw-r--r--Disease Detection in Watermelon Farming Using Robotic Automation[Robotics, ROS, Computer Vision, Python, Published Research]2024-01
Published research (AIP Conference Proceedings, 2024) on a four-wheeled autonomous robot that spots crop disease early in watermelon fields. I led the design of the drivetrain, sensors and ROS software. It follows crop rows by camera-based odometry and fuses GPS and IMU readings with an Extended Kalman Filter.