Work
Selected work.
What 'we build and ship' looks like in practice - across AI in production, quality engineering, and the web. (All anonymized - no client names.)
AI & automation in production
AI that survives real users and real data.
Production tools and proofs of concept where modern AI does real work - not a demo, the thing itself.
Global consumer-audio brand
Release automation for QA test runs
A deterministic, browser-based tool that replaced a manual per-release QA workflow across multiple hardware targets. It integrates with the team's test-management platform over REST, parses structured git-commit annotations to score and prioritise test cases, and creates every per-device test run in a single action.
Outcome
Release setup cut from ~55 to ~10 minutes (~85% less), and untested code can no longer be silently skipped.
Chemical products manufacturer
Technical documents to bilingual product catalog
An admin uploads a Technical Data Sheet PDF; an LLM reads it natively, extracts every structured field - dilution tables, physical properties, safety statements, compliance - generates the Greek/English counterpart, and produces a draft bilingual product page. A built-in LLM audit scores completeness before publish, and permanent slugs feed the QR codes printed on physical packaging.
Outcome
A PDF datasheet becomes a published, SEO-ready bilingual product page in minutes, at ~$0.05-0.08 of inference per product.
Early-stage venture - proof of concept
Voice to structured data, in real time
Automatic multi-language speech-to-text chained with an LLM for zero-shot intent classification, returning typed, schema-valid JSON from free-form - even code-switched - speech in a few seconds end to end. It handles ambiguous phrasing and stateful session context at inference time.
Outcome
No fine-tuning, no rules engine - just transcription and reasoning composed into one pipeline.
Early-stage venture
Synthetic data generation for testing & seeding
A two-tier generator: a seeded, statistically calibrated stochastic tier - reproducible from a seed/config pair, with a persona library and distribution-based sampling - plus an LLM-agent tier that drives full sessions autonomously. Layered guards keep synthetic traffic structurally isolated from production data.
Outcome
Used for regression sweeps, historical data backfill, and demos - all reproducible from a config.
Internal tooling
Price monitoring
Automation that tracks prices across sources and surfaces changes on its own, removing a recurring manual check.
Outcome
Changes are caught without anyone watching by hand.
Research workflow
AI-assisted competitive & market research
A systemized research workflow: multi-source sweeps, claim verification, and synthesized comparison reports that fed directly into product and website decisions.
Quality engineering
Making things work before they ship.
The discipline under everything: define what 'working' means, test against it, and give teams tools they can trust.
Global consumer-audio brand
QA on complex consumer hardware
Years validating connected audio devices and their certification flows for a global hearables brand - the kind of work where 'looks fine in a demo' is not good enough.
Outcome
Devices and certification flows held up to real-world use and certification standards.
QA teams stuck between expensive SaaS and Excel
Offline-first test management suite
A zero-install desktop app that guides testers step by step through manual test plans - pass/fail/skip, failure notes, environment capture - and emits timestamped HTML and JSON reports. A companion CLI imports those results into the team's test-management platform over its API.
Outcome
External contractors run full test sessions with no account, no credentials, and no network - results sync later.
Web & digital