Candidate evaluation time, GDPR-cleared AI scoring platform
~20%
Revenue increase from a data-based sales tool (Teya)
~12 wks
illeszt built solo from first commit to pilot
01 — Selected work
What I shipped, and what it proves.
At work · MNB-supervised payments group · Feb 2026 – now
Product manager and Scrum Master of a greenfield AI lab in a regulated financial environment. Six AI products taken from discovery to production — three of them below. My part: discovery, requirements, compliance, UAT and release, with a team of engineers.
01 · Recruitment
AI recruitment and candidate-scoring platform
AI pre-screens CVs with a reasoned recommendation, analyses interview transcripts and scores a 90-item personality test. AI recommends, HR decides, every step is audited.
~5 months — first interview to production, through every legal and security gate
30 min → <1 min — personality-test scoring
GDPR/DPIA~230 UAT scenariosaudit trail
02 · Call quality
Call transcript analysis for support and sales
Every call over 60 seconds goes through speech-to-text, personal-data masking and LLM evaluation against a versioned indicator register. Team-level dashboards first; individual evaluation sits behind a separate legal gate.
~5,000 calls/month — where manual QA reached only a fraction
108 indicator blocks — across 4 modules
speech-to-textPII maskingEU AI Actgolden dataset
03 · Delivery
AI test-agent ecosystem for software delivery
A nine-phase, AI-first delivery chain with one hard gate: no approved user story, no release. An AI UI tester turns acceptance criteria into scenarios on every pull request and writes its verdict into the PR.
151-point audit — only 15% of the lifecycle was covered at the start
150 acceptance criteria — in the tester's spec, 24 negative
illeszt — from messy orders to an approved order sheet
Private beta · pilot running
AI-powered SaaS for small businesses. Turns free-text email orders and scanned handwritten order sheets into structured line items; a human approves; the daily purchasing summary and delivery notes follow.
„Ne másolgasd. Ne gépeld újra. Csak ellenőrizd.”“Don't copy. Don't retype. Just check.”
Claude Visionhuman-in-the-loopmulti-tenant SaaSpricing
Building has become the easy part: with AI, many people can build, and a lot can be automated. What still decides the outcome is understanding the business problem and the user's or customer's problem — listening actively before proposing anything — and only then building the thing that solves it.
02
Measure before you choose
Compared two OCR providers on six weeks of real handwritten order sheets. Claude Vision read the day-column headers correctly in 12/12 blocks; the alternative lost the grid entirely.
03
Keep the model's job small
The product catalogue deliberately never goes into the prompt; a deterministic lexicon plus fuzzy matching does the matching. Measured ~60 HUF LLM cost per order against a 144 HUF ceiling.
04
Human approves, system prepares
Uncertain rows are highlighted and sorted first; nothing ships without sign-off. The same rule runs the blog pipeline.
05
Compliance is part of the product
GDPR/DPIA, EU AI Act, MNB readiness, audit logging, DPA and terms are written alongside the spec, not after it.
03 — Experience
Ten years, one thread: making complex things shippable.
Feb 2026 – now
Technical Product Manager (AI) · MNB-supervised payments group, Budapest
Greenfield AI lab in a regulated bank: six AI products from discovery to production.
Sep 2023 – Feb 2026
Business Analyst · Fivesky / Morgan Stanley
Engaged 20+ remote stakeholders daily on a data cataloging program; initiated the shift to Kanban; built copilot AI agents.
2022 – 2023
Strategic Project Manager · Teya
Data-based sales tool increased revenues by ~20%; identified EUR 50K operational savings; led a product team of 6 rolling out a payment service app in Hungary.
2020 – 2022
Innovation Coordinator · Robert Bosch
250M HUF prototype project closed on budget and on time; innovation framework for a department of 80.
2017 – 2022
Mechanical Designer · Robert Bosch
11% cost reduction in series PCB production; design guideline used by 2000+ engineers.