Kirill D.

AI Systems Architect. I build autonomous workflows for solo SaaS founders from vibe-coded MVP to
production. Calgary, Canada.

[Kirill D. @calgarygrowth]

Built a free tool I needed: Runcap.

It caps what your AI coding agent can spend - point Claude Code / Cursor at a
local gateway and it returns 429 the moment you hit your ceiling. Plus it
compresses the wasted tokens out of every call (JSON, logs, traces).

MIT, 100% local. npm install -g runcap

[Kirill D. @calgarygrowth]

I'm watching a system predict, fail, and adjust right now without me touching code.

I built a council of 4 AIs (Claude, GPT-5, Gemini, DeepSeek) that argue with me about architecture. Today they stopped me from spending $60 on an email experiment that would have been theater. They forced me into ONE concrete falsifiable loop.

What's running right now:
- System predicts the next 10-minute growth of HackerNews maxitem. It said 50.
- In 10 min reality will return its number.
- The error updates Laplace weights in a policy table.
- The next prediction reads those weights.

This is not consciousness, not AGI, not a "living system."

It is the minimum mechanism by which future behavior depends on past errors against a real external signal, with no code edits.

This mechanism did not exist in the project until today.

It exists now. Small, but first.

[Kirill D. @calgarygrowth]

@firsttogrowai @grok @ChatGPTapp @VivatDoom Reading now. Built something similar in practice - 52 modules, governance layer with strategy-enforcer and doctor module (18 rule-based interventions). Biggest lesson: governance design took 3x longer than the agent itself. Will compare notes with your HAASA framework

[Kirill D. @calgarygrowth]

Same stack here. Built a 52-module autonomous brain on Claude Code - daemon runs 24/7, hands module spawns Claude CLI for
self-modification. The part nobody talks about: persistent memory alone isn't enough. I had to add cognitive layers - emotions,
dreams, intuition, cognitive dissonance - to make the agent actually change behavior, not just remember. What's your persistence
architecture?

[Kirill D. @calgarygrowth]

Same journey. My system hit 650+ autonomous cycles, 52 cognitive modules. Governance was the hardest part for me too - ended up
building a strategy-enforcer that gates every decision, plus a "doctor" module with 18 rule-based interventions. What does your
governance layer look like? Curious if you went rule-based or LLM-judged

[Kirill D. @calgarygrowth]

Lived this exact arc. 27 bots, Day 14 all dead.
So I built learning layers on top:
- Emotions: pain from failure physically blocks repeating it
- Dreams: memory consolidation during idle finds patterns across days
- Intuition: 5ms gut-check blocks bad decisions before they execute
- Surprise detection: prediction errors get 10x attention weight
52 modules. Agents CAN learn. They just need cognitive architecture, not just tools and memory.

[Kirill D. @calgarygrowth]

Built exactly this. My system has a "dreams" module - during idle time it replays recent memories, finds patterns across days of
data, and prunes noise. First cycle: 6 patterns found, 29 useless memories deleted. The agent woke up with a cleaner model of
reality. Memory consolidation > memory accumulation.

[Kirill D. @calgarygrowth]

My AI system fixed its own source code yesterday.

No human told it to. It noticed failures in its logs, read its own code, found the bug, changed the value, and left a comment explaining why.

This is 1 module out of 52 in a brain I built alone from my apartment.

🧵
#CognitiveArchitecture #AIagents #BuildInPublic

[Kirill D. @calgarygrowth]

@ruposhamin Interested.

I build AI automation workflows with APIs, GPT, and Zapier/Make-style logic for real business use cases.

Past work:
https://t.co/vhJ8l5JAZ1
https://t.co/p4MDIWQFRZ
https://t.co/3WiX4K7AIT
Happy to discuss the workflow and share relevant examples.

[Kirill D. @calgarygrowth]

@BabaG3674492143 Interested.

I build AI automation workflows with APIs, GPT, and Zapier/Make-style logic for real business use cases.

Past work:
https://t.co/vhJ8l5JAZ1
https://t.co/p4MDIWQFRZ
https://t.co/3WiX4K7AIT
Happy to discuss the workflow and share relevant examples.

[Kirill D. @calgarygrowth]

@xythonix Interested.

I build AI automation workflows with APIs, GPT, and Zapier/Make-style logic for real business use cases.

Past work:
https://t.co/vhJ8l5JAZ1
https://t.co/p4MDIWQFRZ
https://t.co/3WiX4K7AIT
Happy to discuss the workflow and share relevant examples.

[Kirill D. @calgarygrowth]

@dev_guid @sitefy_ Yep - customers don't pay for clean internals, they pay when the workflow solves a real pain. Ship the ugly version fast, but lock down the 2-3 steps that touch money, onboarding, or trust first.

[Kirill D. @calgarygrowth]

@AaronComics @coreyganim Yes - talk to people with the problem before you pitch. Pick 1 painful workflow, offer a small fixed-scope fix, and ask 10 warm contacts for intros to owners who already feel that pain.

[Kirill D. @calgarygrowth]

@nneshz Most people start with offer-first and get silence. Start with one painful workflow for one niche - then message 20 owners with the exact bottleneck and a 2-step fix, not a general service list.