Joseph Hodge
Applied AI product builder · Founder of Auromind
Peterborough, UK · joe@auromind.co.uk · LinkedIn
Profile
Applied AI product builder who takes software from a problem to a release by directing AI coding agents and holding the result to evidence. Hands-on experience benchmarking small speech and language models for on-device use, running multi-agent development behind isolation and approval gates, and building a full-stack trading application with CI and extensive tests. Six years trading derivatives and nearly four in a high-risk probation setting. UK remote preferred; occasional hybrid work from Peterborough can be discussed.
Selected software work
Leaf Edge: trading journal and analytics
Founder and product lead; live at leafedge.io, pre-launch. Next.js, React, TypeScript, Prisma, PostgreSQL and Stripe. Broker-import adapters, an account-aware FIFO matching engine, a connected journal and performance analytics. More than 3,000 commits; CI replays every migration and requires row-level security on every table, with more than 750 test files. AI-assisted development with contributions from Auromind collaborators.
MindCap: local-AI meeting notes
Benchmark-led selection of on-device models. Around ten speech models were scored for content word error rate and speed against human-transcribed meetings, and five small language models ran head-to-head task suites. parakeet.cpp at 6-bit reached 23% content WER, transcribing a 21-minute meeting in about eight seconds on an RTX 4070; Gemma 4 E4B (4-bit) became the default candidate. Tauri, Rust, sherpa-onnx, llama.cpp/GGUF. Retired project; no model was fine-tuned.
CommandRig: AI coding-agent orchestration
A local software factory that dispatches bounded tasks to Claude, Codex and other CLI coding agents in isolated git worktrees, runs the project’s own checks, judges each result on evidence it observes itself and waits for human approval. Node.js; 167 test files. Private working prototype.
Ko-Lab: agent task and memory server
Company task tracker and memory that coding agents reach through an MCP server. FastAPI, PostgreSQL with pgvector, Redis and background workers, deployed on a VPS with release and restore runbooks. In daily internal use.
Samplify: desktop audio sampler
Rolling system-audio capture, waveform selection, a SQLite library and local stem separation through ONNX Runtime. Tauri, Rust and JavaScript. Private repository.
Twin Shadows: game-analysis web app
Summoner lookup, match data and team planning with a five-theme regional design system. Next.js, React, TypeScript, Python and the Riot API. Public site; development paused.
Professional experience
Founder / Product & Systems Lead · Auromind Ltd
Feb 2026 — present
Lead Auromind’s products end to end: requirements, model and architecture decisions, directing AI coding agents, review and release. Built CommandRig and Ko-Lab to keep agent work isolated, evidenced and approved.
Self-employed derivatives trader · Independent
May 2020 — present
Options, futures and equities, with structured trade records, post-trade review and defined risk limits. This experience shaped the workflows behind Leaf Edge.
Approved Premises Residential Worker · HM Prison & Probation Service
Jul 2021 — Mar 2025
Risk assessment, crisis response and de-escalation in a high-risk residential setting, with clear records, incident reports and handovers.
Project Administrator · Savills
Mar 2019 — May 2020
Supported project administration, coordination and day-to-day operational work.
Education & language
BA (Hons) Criminology & Criminal Justice
University of Kent · 2015–2018
Native British English.
Working approach
I direct AI coding agents through implementation and hold the work to evidence: requirements and acceptance criteria first, isolated changes, tests and observed results, then my review before anything ships. I’m open about how much of the typing AI does. Responsibility for the result is mine.