MUNKANAPLÓ / WORKLOG
Not a case study. An engineering worklog.
Finished systems hide most of the engineering work behind them. This log keeps the problems, wrong turns, fixes, trade-offs and verification steps that do not fit on a portfolio card.
103 of 103 entries are currently available in English. The remaining entries stay in Hungarian while translations are reviewed in batches.
18 entries ezen az pageon · 103 total
An autonomous morning job digest is more than cron plus search
The morning digest became genuinely useful only when search, ranking, CRM state and follow-up work were combined into one idempotent workflow instead of another scheduled list of links.
Open entry →Job searching stays chaotic while the email inbox is the database
A major step for the agent was giving the application process its own CRM ledger, connected to Gmail and Calendar watchers, Telegram notifications and explicit next actions.
Open entry →An onsite US role should not make a developer look less qualified
One ranking flaw appeared when geographic eligibility and professional fit began leaking into the same score. The system needed to answer those two questions separately.
Open entry →Keyword-based job scoring falls in love with the wrong roles too easily
Ranking v2 was born because naive keyword scoring pushed some postings to the top simply for mentioning familiar technologies even when the actual role was a poor fit.
Open entry →The scraper sometimes mistook half a page for a single job description
Some ranking failures were not scoring bugs at all but data-quality problems: aggregator navigation, multiple listings and page chrome had been merged into one supposed job record.
Open entry →The LLM never saw the whole job market — only the uncertain cases left after deterministic filtering
Grounded AI requirement review deliberately ran after deterministic ranking and only on a bounded set of candidates. The LLM complemented the predictable system instead of replacing it.
Open entry →The AI was too eager: it turned responsibilities into must-have requirements
A requirement-review regression showed that the model could promote a responsibility into a mandatory qualification. Explicit evidence guards and a deterministic fallback were needed around it.
Open entry →A local LLM is not slow if you do not cold-start it for every question
Ollama requirement review was paying too much latency for model startup and unnecessary reasoning. Warmup, retry logic and a structured no-thinking fast path made it practical.
Open entry →An 82 match score means nothing if we cannot explain where it came from
While tuning ranking, I built a dedicated diagnostic export and AI-health tooling. The final score was not enough; the system needed to reproduce the inputs and intermediate signals behind the decision.
Open entry →
Analytics plugin updates need verification, or the measurement cannot be trusted
Updating MonsterInsights and Google Site Kit looked like routine WordPress maintenance. In reality, the tracking chain required its own verification because a silent measurement failure can remain unnoticed for days.
Open entry →
When hosting started working against the release process: PogoZone → DigitalOcean
Moving from PogoZone to DigitalOcean was not a simple hosting-provider switch. WordPress, commerce, data, DNS and email dependencies had to be separated while keeping production controllable and recoverable throughout the migration.
Open entry →Email is not an MX record: routing, sending and authentication had to be separated
The hosting migration made email the clearest example of why treating 'mail' as one box is dangerous. Incoming addresses, application-generated transactional mail and domain authentication needed separate lifecycles.
Open entry →Help the player, but do not play the game for them
Guided gameplay needed a careful boundary: explain why an action is invalid and what options are available next, without taking the strategic decision away from the player.
Open entry →A highlight should not infer its board position from a translated label
Board highlights and the latest-track marker became stable only after identification stopped depending on localized text and on behaviour that varied across browsers.
Open entry →One geometry system instead of three separate maps
Country-specific board coordinates were demanding more and more special-case code. We moved to a shared geometry model with validated coordinate data for each variant.
Open entry →We did not leave a bad asset in a 'we will fix it later' state — we reverted it
A supposedly clean USA board background turned out to be incomplete. Instead of layering compensating code on top, we reverted it and rebuilt the artwork around a layered rendering system.
Open entry →Grant's official track scheme became a test instead of a memory
USA tile positions were aligned to the official track scheme, then the important SVG intersections were captured in automated tests so a later layout refactor could not move them silently.
Open entry →USA and Canada should not become two separate games in the codebase
While refining tracks, rounds and the Canadian edition, the goal was to model edition differences as data instead of cloning rule branches for each country.
Open entry →No matching entry in the complete worklog.