Built

I build the product and the way the team ships it.

I start by finding the wider problem beneath the request. Then I stay close to the code and the team until the product works for its users and the business can support it.

selected product work

Farewill

Product & Operations Lead · 2021–22

Probate operations

Led product, operations and workflow design

view the service ↗

We gave each case a structured path and the legal team an operating system.

I worked between customers, legal specialists and operations. We built guided intake that populated case data, automation and audit logs around the legal workflow, and a tracker that showed the next action and helped the team estimate the specialist time a case would need.

I also shipped integrations with HMCTS, HMRC and banking services, reserving the legal team for decisions that needed their judgement.

Agent errors down 69% · case handling from two weeks to four days

Zero Gravity

Founding hire #4 · Head of Product · 2022–26

Learning pathways

Led product and design with the engineering team

see learning at Zero Gravity ↗

Partner-funded learning paths filled the gap.

In 2023 I led the team that shipped staged courses for partners including Accenture, HSBC, KPMG and Snap. They combined video, Duolingo-style progression, quizzes and an early AI skills check.

The product gave students a self-serve layer alongside mentoring and gave commercial partners a concrete way to fund access and preparation.

What we learnedStudents did not return to the library enough. That miss shaped Career Co-pilot: bring the next useful thing to the student instead of waiting for them to browse.

Career Co-pilot

Led product, design and team delivery

hear the podcast ↗

Career Co-pilot turned the catalogue into a guided next step.

I designed and led the first end-to-end AI product on the platform. It used a student’s CV and Zero Gravity profile to build the CV with them, suggest relevant mentors and learning, retrieve useful community posts and answer career questions against platform knowledge.

The CV grew with the student, and the co-pilot brought the next useful part of Zero Gravity into the conversation.

First end-to-end AI product on the Zero Gravity platform

AI STEM tutor

Led product and design · wrote 28% of merged code

view the product ↗App Store ↗

The tutor coaches towards the answer and refuses to hand it over.

Coaching, practice, marking and assignments run as separate agents, each with its own pedagogy and evaluator. We test marking against past papers and official mark schemes, and record safety signals on each interaction.

We launched across Maths, Physics, Chemistry and Biology for AQA, Edexcel, OCR and IB, direct to students and through a school hub for teachers. Eleven weeks later, the government selected us for its AI Tutoring Tools Pioneers Programme: eight companies chosen nationally to test safe AI tutoring in schools. We placed 2nd, scoring ahead of frontier US labs and the largest UK curriculum incumbents.

~67% → 99%+ on internal marking evals · App Store in 45 days

Elliot leaning over a laptop while working with another person — I work through the prototype with the person at the keyboard.

Independent work

My independent work starts with problems I run into myself.

I built Argus because I could not retrieve my research, and crux because a code diff could not explain the decisions behind it. I apply the same test to each project below: use it in my own work and keep changing it until I trust it.

argus

My private research system for product work, career questions, startup ideas and LLMs. It takes transcripts, feeds and my own notes, then builds evidence-backed views I can question or correct. Argus measures useful shifts in a view rather than the size of its corpus.

The cheap jobs are code: fetching captions, deduplicating sources, validating paths and rebuilding views. The model gets a compact excerpt and one decision to make. It keeps, discards or holds the source for review, then says which existing view should move.

signals[input]
raw evidence
Transcripts, feeds, papers, drafts and my corrections.
ingest[code]
fetch + clean
Fetch, reject junk, deduplicate and validate every path.
triage[LLM]
one decision
Keep, discard or hold; then name the view that should move.
views[code + me]
current model
Evidence, implications and every before-and-after stay visible.
answers[ask]
bounded retrieval
A useful answer, cited back to the material that supports it.
me[decide]
human judgement
Question the answer, correct the position or make the call.

correction.patch → views

The constraint The model judges what is worth keeping and where it belongs. Code does the fetching, filing and validation. I question the output, and a correction changes the view—not its history.

research

crux

You are shipping faster than ever. Are you getting sharper, or just getting carried? Nothing currently measures that. I noticed it in myself at Zero Gravity: shipping faster than I ever had, and slower to say what I would have done differently. Output has never been higher and no instrument tells you whether the person behind it is improving, plateauing or quietly atrophying. Fluency frameworks answer the baseline and everyone will pass them; the layer above is where the difference sits, in trust calibration, resistance to output that looks polished, and knowing what to kill.

crux measures it. A Claude Code hook reads each session and extracts what the human actually decided: what got rejected, redirected or killed while the model did the typing. Ongoing research rather than a product, published with the method, the results run on myself, the honest objections, the limitations, and a memo to the platform layer about the half that nothing measures.

read the research · repo

agent tools

The agentic vocabulary, mapped to what is above: orchestration and multi-agent patterns in the tutor and argus, tool use in the MCP servers and skills below, retrieval in Career Co-pilot, human oversight and evaluation on /loops and /evals. Memory is thin and planning is absent: I have designed stopping rules, not planners.

boulot

A free, local career system that keeps your experience, applications and outcomes on your own machine. It learns which claims earn a reply and carries that evidence into the next application.

last pushed 2026-08-03

claude-skill-potions

Curated Claude Code skills for ops and product workflows. The skills directory is 40k+ deep; these are the ones that actually work.

★ 62 · last pushed 2026-02-12

vox

Voice of Customer research agent. Eight days of PM research in eight minutes: JTBD, personas, opportunity mapping from Gong, Granola and Jiminy data.

dabble

Visual editor for server-rendered (Hotwire) apps. Edit the running app in place, write real ERB. Kills the design-to-code handoff for the stacks React-first tools ignore.

ward

A safeguarding layer for LLM apps that serve under-18s. It screens each message for a genuine disclosure, separates that from ordinary bad conduct, and routes the real ones to a named human on a clock. Grounded in KCSIE rather than generic content moderation. Built around the precision problem: page a Designated Safeguarding Lead on every false alarm and they stop trusting the alerts, which is worse than having none. On its published synthetic eval sets, the Claude judge reaches 90% recall at 100% precision and a 0% false-positive rate, against a keyword baseline at 50/83/8.6.

last pushed 2026-07-25

homebuyer-mcp

UK home-buying MCP server. Conveyancers and mortgage brokers from live SRA, FCA and Companies House registers, plus stamp duty, lease checks, survey explainers and title register analysis. Eleven tools.

last pushed 2026-07-29

hooksmith

Browse and install pre-built Claude Code hooks with one command. Twelve hooks, zero config. The missing package manager for hooks.

ideas or feedback?

Two systems here are real but private: argus, an agent fleet that turns research into working views, and LifeOS, the front door that routes my whole setup. Ask me about either: elliotjlittle@gmail.com.