Tiny models under hard limits. Web that actually ships.
Not a pitch deck — a lab log you can fork and argue with.

portfolio · parameter-golf writeup

parameter-golf / HYDRA — fork of openai/parameter-golf

Stack: PyTorch · quant · schedules · multi-seed checks.
Notes + runs: writeup

Also on the board? Open an issue on the fork — compare runs, no ego.

ZeroResp v5.2 — adaptive Iterated Prisoner's Dilemma player for Axelrod-Python. Field pack (32 strategies, 200 turns, 8 reps): 2.843 SPT at 0% noise, 2.375 at 5%. Reciprocators / Grudger hit 602. Production freeze — remaining oracle “gaps” are exploiter ceilings, not bugs.

  16MB / 10min ──► squeeze the model ──► measure BPB ──► do it again
         ▲                                           │
         └──────────── receipts or it didn't happen ─┘

Stuff you can open in a tab — not “coming soon”.

Side experiments (CSS / layout playgrounds): Teleboxi · apikeis

Pins (profile UI): ZeroResp → parameter-golf → EchoPrint-AI → Evreu1pro.github.io

1. One hard problem — parameter-golf: limits you can measure, not vibes.
2. One sharp lab — EchoPrint: what the browser leaks, scored in public.
3. One game-theory player — ZeroResp: tournament receipts, then stop at the ceiling.

No “open to work” banner. No badge farming.
If something here is useful — star it, fork it, roast the writeup, send a PR.

• Langs: Python · TypeScript / JS · HTML/CSS when zero-build wins
• Ship: Vercel for web · local receipts for ML
• Style: design tokens, product-feeling static when it fits
• Git: small PRs when I can; side projects get honest messy commits
• AI: yes, as a tool. Decisions and late-night fixes are mine — check history if a file looks “too clean”

Chat: RU / EN / DE · code comments: usually EN

Small clear docs/fix PRs welcome. Empty ⭐ spam — not the game.

Evreu1pro · Leipzig · public lab · 2026
receipts > aesthetics · limits > slogans