PABLO.ENGINEERING
GITHUB
WORK & LEARNING LOG: ACTIVE
// WARSAW, PL ·

I BUILD AI AGENTS.
I LEARN BY BUILDING.
I WRITE DOWN WHAT WORKED AND WHAT DIDN'T.

This is my log of work and learning. I write Python and TypeScript, build AI agents for my own use, and write down what works, what breaks, and what I learn along the way. A lot of code is written with LLM help, so I try to test and read it before it goes further.

EXPLORE PROJECTSREAD THE DEVLOG
TELEMETRY // ORCHESTRATOR
v2.4 // SNAPSHOT
CORE RUNTIMENode 22 LTS / Python 3.12
ACTIVE AGENTS8 Specialized Roles
STORAGE & QUEUESQLite (WAL) + Event-Driven
DIGEST ENGINEAutomated Git Pipeline
Daily Auto-Digest: Git commits are synthesized into devlog articles every 24 hours.
// 01. LAB PROJECTS

SELECTED PROJECTS

[ PROJECTS & EXPERIMENTS ]
CORE SYSTEM

AGENT OPERATIONS CENTER (AOC)

A panel where I manage eight AI bots: PM, developers, QA, and security. I start tasks in the background with `child.unref()` in Node.js instead of polling the database in a loop. Writes go through SQLite in WAL mode, so I don't need Redis. The whole thing runs on one small VPS. How many tasks it can handle is something I'm still checking.

STATUS
IN DAILY USE
ARCHITECTURE
EVENT-DRIVEN
SPECIALIZATIONS
8 ROLES
#Next.js 15#TypeScript#Zod#SQLite WAL#Event Broker#SSE Streaming#Docker

APPSEC & QUALITY GATE

A set of rules for code that AI agents write: no TypeScript type bypasses (`as any`, double assertions), input validation with Zod, and a separate reviewer agent that checks every commit. This doesn't guarantee bug-free code. It's a filter that catches the simplest mistakes before they reach the repository.

[AUTO-AUDIT] DAILY VIA CI
  • Anti-Slop TS Guardrails
  • SAST AST Inspection
  • Strict XSS Sanitization
// AI AGENTS // GARMIN .TCX TELEMETRY

AI COACHES (TRAINANBOT & PABLOTRAINER)

TELEGRAM LIVE

Two Telegram bots: TrainAnBot and PabloTrainer. They read Garmin `.tcx` files and calculate heart-rate zones, so I don't enter workouts by hand. TrainAnBot draws on PubMed publications, while PabloTrainer watches training load across boxing, gym, and running. It's an experiment, not medical advice.

#AI Agents#OpenClaw#Telegram Bot#Garmin .TCX#PubMed API#Combat Conditioning
// AI AGENT // CODED PEDAGOGY

PROFECLAW - SPANISH TUTOR

A Telegram bot I use to learn Spanish. It corrects at most two mistakes per message so it doesn't discourage me, tracks repeated errors, and builds review sessions from them. An eval suite checks whether the bot invents etymologies. That reduces those errors but doesn't eliminate them.

#AI Agent#OpenClaw#Spaced Repetition#Eval Suite
// R&D LAB

DEEP CS & VIBECODING LAB

This is where I practice algorithms and data structures in Python and TypeScript, harden a Linux server, and build small tools. Git commits turn into a draft post once a day: two Gemini models, one writes, the other checks.

#Algorithms#Python#TypeScript#Linux
// 03. ARCHITECTURE & CAPABILITIES

ENGINEERING TIMELINE

[ TIMELINE & STACK ]

I built a panel where I manage eight AI bots. Instead of polling the database every few seconds, I start tasks in the background with `child.unref()` in Node.js. Task start time dropped from about 4 seconds to 50 ms (described in the 16 August devlog). SQLite in WAL mode allows non-blocking writes, without costly clusters like Redis. I'm still checking how the system behaves with more tasks.

Next.jsTypeScriptSQLite (WAL)Event-DrivenDockerSubagents
// 04. LAB / CONTACT

SOMETHING INTERESTING?
WRITE TO ME.

Got a problem that fits AI agents, event-driven architecture, data analysis, or unreliable automation? Write to me. I reply to every sensible message.

contact@paterski.com
GITHUB PROFILE