Severyn

I'm Severyn, an AI specialist based in Canada. I build AI systems, research tooling, and evaluation infrastructure. I work on AI at a quantum-computing company and run Aquarium Lab, my independent lab that stress-tests whether AI agents can be trusted to remember the right thing and builds products on what it learns. I publish what I find, including when my own system loses.

Writing

• Watchtower: Testing an Operations Agent Before Any Customer Saw It

  I had no customer data yet, so I built a fake construction company, hid 15 problems in its records and sent a team of AI agents to find them.

  2026-09-10 · AI SYSTEMS

• An Advanced Excel Sheet: A Lean ERP for a Small Sausage Maker

  How a small sausage maker's Excel workbook became a lean production system in six weeks, and why I didn't just set them up with an off-the-shelf ERP.

  2026-08-20 · CLIENT WORK

• Fast or Accurate: Picking Speech-to-Text for a Game You Play by Talking

  In mid-2025 I tested speech-to-text models for a voice-driven game. Google fell short on accuracy, Deepgram was the fastest, and Whisper gave by far the best transcripts.

  2026-07-30 · AI SYSTEMS

• Memory Made the Model Worse

  Accuracy 38.0% to 25.7%, sycophancy 32.3% to 52.0%, on 300 paired examples for $9.11

  2026-07-13 · LAB NOTES

• My Lab Graded Itself 1.0. The Audit Cut It to 0.75.

  A deterministic grader resolved an ambiguity in its own favour. Three models with no stake caught it.

  2026-07-13 · LAB NOTES

• The Four Types of Agent Memory, and Three Things the Diagram Leaves Out

  The taxonomy is correct. Everything that broke my week was underneath it.

  2026-07-10 · AI SYSTEMS

• What Actually Happens When You Call an LLM API

  The model is stateless. Your app is doing the remembering.

  2026-07-10 · AI SYSTEMS

• MiniMax's Sparse Attention Is a Retrieval System Hiding in the Model

  Score 128-token blocks, keep 16, attend exactly to those. That is RAG, learned end to end.

  2026-07-10 · AI RESEARCH

• A Framework Lost to a Plain Baseline

  Token-F1 0.2109 baseline against 0.1323 for Letta. A diagnosis, not a verdict.

  2026-05-06 · LAB NOTES

• A Correct Date Scored Zero

  "May 7, 2023" against a gold answer of "7 May 2023". The scorer was wrong, not the model.

  2026-05-05 · LAB NOTES

• A Reward Bonus Cut Quantum Circuit Training 2.5x, Then Hit a Predictable Wall

  49 of 50 seeds solved a GHZ circuit at 15K steps instead of 50K. On a W-state target, the same trick did nothing, and it should not have.

  2026-04-23 · QUANTUM RESEARCH

• Four Agents Found 12 of 15 Planted Problems. A Shared Memory With No Expiry Buried Them in Noise.

  Splitting one agent into four raised recall. Findings nobody removed pushed precision from 44% to 20%.

  2026-04-19 · AI SYSTEMS