AI Engineer in Calgary, 8 years of production backend behind it.
88 seconds on two things I built:
• JobPilot AI: multi-agent job pipeline, 5 LLM providers, 1,250+ tests
• go-mcp-bridge: MCP with human consent + scoped tokens
https://t.co/OErtKdub6S · DMs open
Fine-tuning changes how a model behaves, not what it knows.
I made a 5-min explainer on when it actually wins, why labels and bigger teachers can hurt, and why you build the eval first.
Sources are on screen in every scene.
#LLM#FineTuning#AI#MachineLearning#LoRA
🎬 New 12-minute explainer on building verification infrastructure for coding agents.
Four layers, cheapest first:
1️⃣ Static constraints. The architecture rule lives in lint, so a violation fails CI with an error that names the fix.
2️⃣ Executable tests. Red first, with expected values taken from the requirement. Mutate slugify to return undefined and the strong test goes red while the weak one stays green.
3️⃣ An independent reviewer. The agent that writes the code never reviews it, and green CI doesn't count as a verdict.
4️⃣ Runtime evidence bound to a commit. A rebase moves the SHA, so the old verdict goes stale.
After that: a fixed CLI for doctor, drive, evidence and cleanup, the feature map, my 0 to 4 trust ladder for parallel agents, and six failure modes, including Ray Fernando burning over two billion tokens in a day.
Built from pstack in cursor/plugins and Lauren Tan's (@poteto) talk. Music only, captions on screen.
#AIAgents #SoftwareEngineering #DevTools
Coding agents say "done" every five minutes. I wanted to know what would make that word mean something.
The answer I keep coming back to is a verification stack, so I drew it as a live panel that runs like a monitoring screen: L1 constraints in lint, L2 a failing test written first, L3 review by an agent that didn't write the code, L4 evidence from running the real product.
Then the loop that repeats under it: launch, doctor, drive the real user path, keep evidence, clean up. Any check fails, run it again.
34 seconds, no narration, captions included. Diagram generated from JSON with the open-source live-panel engine; every claim on screen comes from pstack's create-verification-skill and Lauren Tan's talk on shipping with agents.
#AIAgents #SoftwareEngineering
Honestly, I don't think using CLI is an ego thing. It depends on users' preference. CLI used to be better than APP, but Claude Code has put a lot of effort into improving it. The user experience, I guess, right now, there is no such big gap; it's totally your preference and your habit, nothing more.
Still checking 5 websites for AAIP draw scores?
https://t.co/gfUYNm7WLC puts Alberta draw cutoffs, timeline estimates and an AI assistant that cites its sources in one place. Free, in EN / 中文 / FR.
Based on official public data · Not a government agency · Not legal advice
#CanadaImmigration #AAIP #ExpressEntry
try pstack! it’s my personal set of skills for rigorous and autonomous engineering with agents. it’s updated frequently as I continue to tune and simplify them
10 second tutorial:
1. run /setup-pstack and choose your budget
2. add /poteto-mode in front of any prompt you want done rigorously
3. ask /poteto-help whenever you’re unsure what pstack skills to use
4. that’s it!
Grok Bot: https://t.co/nqeRdY62ee
Cursor: https://t.co/WDB4U1rwmu
Source: https://t.co/u9xIlN3eN0