Eidetic

Robin Miklinski — software systems & applied AI · London

I build AI systems, and the controls that decide what they can do without asking.

Fifteen years across defence, retail, collaboration software and global SaaS. Three of those companies were acquired, one for £1.15bn.

I need something built → I’m hiring → Show me the work →
  1. Question how the operation works today
  2. Explore plausible ways to improve it
  3. Remove what adds no value
  4. Build the smallest complete system
01 / Ledger

Fifteen years of the same instinct.

None of this was built for a website. It is how I was already working.

Seven public repositories, linked below. Three of them tell an AI what it must not do, in a file that ships with the code.

Year Context / System The Gate / Irreversible Constraint Access
2026 producer-release-os AGENTS.md: no backend, no auth, no payments in v1. No publishing or deploying without asking. Public repo →
2026 eidetic-studio Never fabricate a button sequence. Unverified steps marked and sourced. The knowledge base is the authority. Public repo →
2026 yr2l-emails Draft-only by design. Every reply needs human review. Public repo →
2026 ai-complaint-resolution Drafts FCA complaint responses. The QA report is published with the code: 180 test cases, plus cost guards in CI. Public repo →
2026 AI Go/No-Go Gate (Reward Gateway) Refused a binary sign-off. A panel accepted the risk in writing instead, and the evaluation suite became the gate for every AI launch. Narrative Case
2025 eidetic-sample-tools Non-destructive export by default. The library is more valuable than the software. Public repo →
2025 Agent Infrastructure (Mac mini) A linter that blocks bad writing and remembers the rule. 95 operations allowed, 30 denied outright. Private Architecture
2024 BrowserStack World Tour & QE Keynote on AI testing, negative-test blitzes and building the capability inside a team rather than buying it. Speaking
2023 Reward Gateway Exit (£1.15bn) Built the quality and evaluation architecture in the run-up to the £1.15bn Edenred acquisition. Public Record
2021 cypress-slack-reporter Open source CI reporter for Cypress test results directly to Slack and CircleCI. Public repo →
2019 ZAPClient OWASP ZAP security client in CI pipeline. Automated vulnerability scanning. Public repo →
2018 Tooling Budget Consolidation Left the vendor and rebuilt the tooling in house. A six-figure saving, and the capability went up rather than down. Commercial
2016 Black Friday Peak Trading Load Five-figure concurrency simulated ahead of peak. No performance incidents on the trading days themselves. Performance
2015 OWASP ZAP Security Tests (Huddle) Security scanning moved into the TeamCity pipeline, so a scan ran on every build rather than before a release. CI Security
2010–14 Defence Data Systems Threat-modelled data systems where an irreversible action had to clear a gate before it ran. Defence
02 / The Story

The AI Go/No-Go Gate: Informed rather than surprised.

An AI product assistant was ready to ship and I was asked to sign it off. One approval, yes or no. I refused. The same prompt does not return the same answer twice, so a pass mark on the day I tested it would have said nothing about the week after.

I built an evaluation suite instead: adversarial prompt injection, hallucination bounds, sensitivity checks. It does not return a verdict. It shows how the assistant fails and where, which is what a launch decision actually needs.

The decision went to a panel who accepted the risk in writing, with the failure modes in front of them. I instrumented follow-up question rate as the product metric, on the reasoning that a user asking again is a user who was not answered. That suite is still the gate every AI feature there goes through.

03 / Workbench

Public repositories, and the constraint each one was built under.

ai-complaint-resolution Python

Complete AI product with published QA report

Multilingual response drafting for FCA-regulated complaints. AI proposes, human approves: that is a feature, not a setting. The 180-case QA report is published with the code.

Applied AI View Repository →
producer-release-os TypeScript / React 18

Local-first planning tool for music producers

Token-driven React 18 SPA consuming a custom component library. Explicit refusal in AGENTS.md: no backend, no auth, no payments in v1.

Interfaces View Repository →
eidetic-sample-tools Python

Sample library curation for hardware samplers

Non-destructive exporter for Octatrack MKII, Digitakt and TR-8S. The library is more valuable than the software, so nothing is ever written over.

Services View Repository →
eidetic-studio Architecture

Git-tracked source of truth for physical studio

MIDI clock chain, decision records and capture protocol. The agent may not invent a button sequence: anything unverified is marked as such and sourced.

Direction View Repository →
cypress-slack-reporter TypeScript

Published CI reporter for test results

Open source reporter that puts Cypress run results straight into Slack and CircleCI, so a red build is visible where the team already is.

Platform View Repository →
ZAPClient Python

OWASP ZAP security scanning client in CI

Automated security scanning wrapped for continuous integration pipelines. Committed in 2019, when running a scan on every build was still unusual.

Platform / Security View Repository →
04 / Measurement

Measurement over decoration.

I changed my own headline metric’s definition halfway through the year, because the old baseline was not defensible, and wrote down why. The number got worse. The measurement got real.

Every metric I publish ships with the rules that produced it, so anyone can check the arithmetic instead of trusting the dashboard.

05 / Capability Index

  1. 01 Interfaces

    • workflows
    • integrations
    • design systems
    • React 18
    • token architecture
    Evidence: producer-release-os →

  2. 02 Services

    • audio processing
    • APIs
    • metadata cleaning
    • background tasks
    Evidence: MP3-Service → · yr2l-emails →

  3. 03 Applied AI

    • retrieval
    • classification
    • agents
    • evals
    • guardrails
    • red-teaming
    • prompt injection
    Evidence: ai-complaint-resolution →

  4. 04 Platform

    • architecture
    • reliability
    • performance
    • CI/CD
    • security scanning
    Evidence: cypress-slack-reporter → · ZAPClient →

  5. 05 Direction

    • technical strategy
    • feasibility
    • decision records
    • risk governance
    Evidence: eidetic-studio →
06 / About

Robin Miklinski

Background

For fifteen years I have built systems across defence, retail, collaboration software and global SaaS. I read Internet Engineering at Exeter and started out in defence data systems, where the boundary around what a system was permitted to do was the first thing anyone drew.

I still write the code. Python, JavaScript, TypeScript and C#, the architecture around them, and the guardrails that decide when a system has to stop and ask.

Outside Work

I produce electronic music and DJ under the name Eidetic. The hardware studio runs on the same rules as the code: capture protocols, decision records kept in git, and an agent that is not allowed to invent a button sequence it has not checked.

07 / Contact

Work together.

Consulting / Review

Work with me

Two weeks on one AI feature or workflow. You get a written findings document, a threat model and an evaluation test plan you can run yourself after I have gone.

  • Evaluation and guardrail audit of an AI feature
  • A testing strategy for output that is never identical twice
  • Cost and rate guards in CI
Book a Reliability Review →
Leadership / Engineering

Hire me

Founding engineer or early technical lead, hands on the code. For teams shipping AI products that have to be right as well as quick.

  • Python, TypeScript, React 18, C#, SQL
  • LLM evaluation, RAG security, PyTest, Docker
  • Technical strategy signed off at CTO and CPO level
Get in Touch → GitHub / @you1anna →