I take an operation that already runs — creative, marketing, business — and put AI inside it where it pays: a pipeline, an approval gate a named person signs, a ledger an auditor can read. Eleven years inside a regulated fintech, then delivery for EU institutions. Based in Bulgaria, EU work rights, no sponsorship needed.
On Snapdragon 7 Gen 4 the NPU returned garbage in llama.cpp: its matrix unit has no FP16 and the software assumed it did. I reframed the bug as the opportunity and, working with an AI coding agent, moved the math to integers. Prompt processing is now 3.3× faster than the CPU and 2.3× faster than the GPU, with the same perplexity and byte-identical answers. A 21-chip map shows which Snapdragons need it. The case page shows the reasoning step by step, dead ends included.
An n8n module for the deployer side of Article 50: classification, burnt-in label, a human gate a named person signs, and an append-only hash-chained ledger in Postgres. The auditor's page is behind header auth and verifies the chain on every render.
Ask a folder of markdown notes a question and every claim comes back with the id of the note it came from, or a plain NOT_IN_NOTES. Hybrid BM25 and bge-m3 retrieval fused by RRF, an evidence gate that stops off-topic questions before any model runs, and a check that every citation was in the context actually sent. Measured on 28 questions with traps: 19 of 19 answered with the right note cited, no invented citations, 8 of 9 traps refused. The one miss is written up, not hidden.
A node canvas for pitching a system before it exists: drag, rewire, animate the flow, fly between camera views. It imports a live n8n instance, so the diagram in the meeting is generated from the workflow, not drawn to look like one.
A consultancy staffs tenders from an archive nobody can read. This turns it into islands you can see, then answers a requirement list in plain language or straight from the terms of reference — every requirement backed by a quote from the CV, re-read by a second model. Vectors on the machine, no data leaving the building.
A live tool scoring tokens in real time on buy pressure and volume velocity, zero dependencies against the GeckoTerminal API. The signal is built on 300 early-bird wallets and validated by bootstrapping over a thousand iterations.
An independent read of how a DeFi analytics platform presents data: signal-to-noise, information hierarchy, what the first view has to answer. Delivered as an authored walkthrough ordered by impact against effort, not as a list of opinions.
The same work, older tools. A process nobody wanted to do by hand, turned into something that runs.
Every one of these arrived with an approval chain, a rights holder and a deadline. What made the work repeatable was not taste. Brief to variants to localisation to upload, the manual step removed through the Figma API, the winning variant picked by SQL rather than by opinion, the result measured in CPA and conversion.
Bulgarian residence, EU work rights, no sponsorship needed. Remote from the EU, contract or employment, English or Russian.