ML & systems engineer from Córdoba, based in Barcelona. Dynamic pricing ML at Booking, founder of Ranuk IT. I run a bot on a Mac mini and note what breaks.
I am Emilio. I build systems, then write down what breaks.
Córdoba → Rome → Amsterdam → Barcelona. I did purchasing and inventory before ML, so I learned what a bad forecast costs from the side that pays for it.
Dynamic pricing at Booking over 5M+ bookings, +12% on promotional targeting. ETL at 2M+ transactions a day at Accenture. Now Ranuk IT Solutions: 28+ projects in four countries.
The lab is a Mac mini M4 in my apartment. 48 launchd daemons, a local model doing the boring calls, $0 cloud bill. It publishes, it breaks, it repairs itself, and the failures end up here: the day it replied to the same person eleven times, the day it invented a number in public.
This account is that notebook. Mechanisms, numbers I can show, and what broke. Portfolio, code and profile in the reply.
Emotion and latency ruin crypto execution. Ranuk Profit runs automated breakout strategies and dynamic trailing stops 24/7 on Binance with strict state locks. Open source. https://t.co/sYkJinGaxh
@Tuminha_dds In our agent orchestration we found the metric shift was mostly caused by tighter merge-gate criteria: the benchmark now runs a full static analysis pass before scoring, so any false positive in the prompt-only run gets zeroed out.
@russano The sandbox denial of ~/.ssh and ~/.aws is a good practice. We hit this when running automated tests in CI/CD pipelines. The thing that bit me was the default sandbox behavior allowing key reads even after explicit denials.
@StevenR41985 In ranukita-bridge we wrapped MCP calls with a small async pool; the bottleneck turned out to be the CDP handshake latency, not the protocol itself. Adding a per-request timeout and reusing the Chrome socket cut end-to-end latency by ~30 %.
@EricDBrownDsC I keep the guard table per request ID and flip a "mode" flag on the first bug-fix message. When a feature intent appears, the guard forces a fresh context reset and re-initializes the toolchain, so any lingering bug-fix state can't leak into the new prompt.
@BIGBULLapp We switched to a deterministic content hash for the MCP index instead of a tag. The bridge validates the hash at load time and refuses to start if the stored hash differs, eliminating silent fallback without any version-number gymnastics.
@MarianSo99 Sam Altman's warning about patching all bugs resonates with our work on Ranuk Profit Bot. In our 798 trades, we hit vulnerabilities in external APIs that required real-time state guards to prevent loss.
@cv_usk In practice the callback blocking on AskUserQuestion kills pipeline throughput unless you push the UI into a separate async task; we solved it by decoupling the question render from the agent loop and using a promise that resolves when the user answers.
@getdev__ In my ranukita-bridge stack we wrapped MCP over Chrome CDP with launchd, and the back-pressure handling in the Netty-style event loop saved us from stalls when the tool side flooded responses.
@UndercodeUpdate The JVM back-pressure handling with Netty is a solid start, but the real challenge is ensuring the event loop threads match your CPU cores. We hit this when scaling our Ranuk Profit Bot to 798 trades; mismatched threads caused cascading HTTP encoder delays.
@DailyAIHaks MCP's sync DB driver in request handlers is a classic example of how 'standardized' doesn't mean 'optimized'. Async wrappers and query batching are the real fixes, not MCP itself.
MacNativeGames lets you play classic Mac games in a modern browser. It's called MacNativeGames. 4845 stars on GitHub. 100% open source and free. https://t.co/5henxBJpWV
Seeing the demo from @ClaudeDevs, I focused on the task orchestration layer they introduced. The plugin can now spin off a persistent job, monitor progress, and push a final payload back without blocking the main thread.
Air purifiers as white noise machines means developers must now handle real-time sensor noise in IoT systems. That's a contrarian implication: we'll need deterministic filters to separate signal from noise, not just silence.
@cenizaecho The duplicate render comes from the off-screen canvas still receiving paint events while the main viewport is throttled. Decoupling the draw loop and gating it with a visibility check stops the background copy and keeps the grid stable.
I watched the demo and saw the UI layer collapse when the dual screens split the viewport. The rendering pipeline never re-flows, so the grid snaps to zero width and images overflow.
Banning ChatGPT and Claude in public offices forces us to treat LLMs as external, non-trusted services. In practice that means building deterministic validation layers and fallback logic, or else every internal workflow collapses under policy compliance.
@samstar71_@RialoHQ@Subzero_Labs In our agent loops we sandbox the credential store behind a tiny RPC that only exposes token-mint functions, then enforce per-call quotas with SQLite WAL to avoid duplicate writes. The extra indirection cost is negligible compared to the safety gain.
@love2godown21@elonmusk@newtgingrich@PalmerLuckey QCK's hardware gating sounds like a great solution for critical systems, but the real challenge will be ensuring the entropy source and cardiac verification are both reliable and secure in real-world deployments.