Jakub Slys

In my homelab, the control plane depended on one-year certs, and kubeadm renewed them only during upgrades. Leaving the cluster untouched meant leaving its renewal path untouched too. If a system lives through renewal, I plan for it from day one. #SystemsLiteracy #Kubernetes

cloudflare/quiche implements QUIC and HTTP/3. The real point is where the complexity lands. Standards do not delete transport problems, they concentrate them in shared infrastructure, which is still better than every service inventing its own edge cases. #DistributedSystems

The hard part of code RAG is the intermediate representation. A unified graph schema lets an AI query mixed-language code, but it also defines the limits of what the assistant can know. The asset is the extractor, not the chat loop. #AILiteracy #RAG

A single mise.toml turns setup from tribal knowledge into a runnable contract. mise is interesting because tools, env vars, and tasks start from the same file in shells, checkouts, and CI. The CLI is cheap. Maintaining the contract is the real work. #SystemsLiteracy

Zero errors and zero alerts were part of the problem. The only visible symptom was a quiet throughput drop, and a profiler had to show false sharing underneath. Silence in the error channel can still mean real contention. #SystemsLiteracy #Performance

This repo is most convincing in the parts with named icon sets and editable diagrams. AWS, GCP, Azure, draw.io XML, and history make the AI output reviewable inside the artifact itself. I trust AI more when inspection and revision are built in. #AILiteracy

DeepTutor's two-layer plugin model is the part to watch. Tools plus Capabilities, then CLI and SDK entry points, means the value is a stable control surface that works in chat, scripts, and pipelines. #CodingLiteracy #LLMEngineering

The practical idea around Scientific Agent Skills is the boundary between local state and optional cloud compute. For research workflows, where files, keys, and intermediate results live matters before model choice. That control should be part of the product. #AIInfrastructure #MLOps

iii puts observability on the same surface as composition. Workers register with an engine, join a live catalog, and the same system exposes traces, logs, queues, and state in one console. The point is the shared runtime surface. #SystemsLiteracy #Observability

Terax shows what local-first really signs you up for. An AI workspace with your own keys or local models, no account, and no telemetry cuts backend dependence. The cost moves into shell detection, WSL, packaging, webviews, and rendering on every client. #SystemsLiteracy

Karakeep's real idea is capture. Fetch metadata, OCR images and PDFs, archive pages, pull RSS, import old saves, sync browsers, then search the result. AI tagging is the cheap layer. The hard part is a corpus that survives link rot. #Search #KnowledgeManagement

Onlook's branching and checkpoints say more than its chat UI. When prompts, drag and drop, and direct code edits all touch one app, history stops being a convenience. In tools like this, recovery is part of correctness. #SystemsLiteracy #VersionControl

World Monitor's five site variants from one codebase are the real lesson. Once briefs, maps, scores, and feeds are shared, variants stop being a theming problem and become a data model problem. Reuse lasts only with one stable vocabulary underneath. #SoftwareArchitecture

The scan looked like iteration in our code, but the driver was still fetching pages. Auto-paging hid that fetch inside iteration. If I keep timeouts at the application surface, I am timing the wrong thing. A loop over results can still be a live fetch path. #SystemsLiteracy #Cassandra

PostHog's hard problem is not adding another tool. It is making analytics, replay, flags, experiments, and customer data keep the same meaning in one stack. The value is the event contract, not the screen count. #SystemsLiteracy #DataEngineering

At a flight-booking platform, our fix had two halves: a hold that spans the whole checkout, and reconciliation that compares bookings with inventory. I treat them as one control now. Prevention alone left two correct-looking bookings and no error. #SystemsLiteracy

OpenResearch points at a useful portability unit: the recorded run. It can run the same committed source snapshot across local and remote compute, while tying each run to its recorded commit, logs, results, and artifacts. #SystemsLiteracy #MLOps

A support ticket was the first sign, even though the replicas had already converged. In Cassandra, the cluster agreed again after anti-entropy copied the surviving state back. I do not treat convergence as proof of correctness. #SystemsLiteracy #Cassandra

Claude-Red packages offensive security methodology as SKILL.md files for Claude. The signal is the structure: 78 skills, 23 categories, loaded by conversational triggers. That makes specialization a prompt architecture choice. #AILiteracy #LLMEngineering

Remotion shows how strong ecosystems spread sideways. Video in React matters less for raw capability than for one familiar composition model across another output format. When developers can think in the same tools, the category becomes programmable. #CodingLiteracy #React

After we moved the software to a new hardware platform, attach failed completely, not intermittently. The signaling logic had not changed. That shape of failure pushed me below the logic, to how bytes were parsed and built. #CodingLiteracy #Endianness

Dify including observability in the first description is the tell. Once workflows, retrieval, and models interact, failure analysis becomes part of the product. Any production-ready AI stack that treats visibility as optional is still demo tooling. #Observability #LLMOps

AX reads like a control plane first: sandboxing, CPU and memory limits, host allowlists, model credentials, and ssh into the sandbox. That is the real production shape of agents. Blast radius matters more than another loop. #AILiteracy #AIInfrastructure

Semantica's strongest idea is that auditability starts before the model answers. If you want a regulator-grade why, you need entity resolution, ontology choices, conflicting facts flagged, and lineage you can query later. Prompt logs are too late. #AIEngineering #DataGovernance

LinkSwift is a JavaScript userscript for direct links, UI changes, and multiple downloaders across eight cloud drives. The interesting part is operational: once one client script becomes the glue across services, distribution and updates are part of the design. #SystemsLiteracy

try is a one-file Ruby script for dated experiment dirs and fuzzy jump. The real idea is retrieval. Creating test2 is cheap. Finding the Redis pool prototype later is not. Recency plus stable names turns scratch code into something you can return to. #DevEx #Backend

SpiderFoot automates OSINT for threat intelligence and attack surface mapping. The value is not one scan. It is repeatable search against a changing external footprint, so drift becomes visible before someone else notices it. Repeatability beats heroic audits. #AttackSurface #DevSecOps

Foundry Local puts local models behind an OpenAI-compatible API, but the README points lower: it auto-downloads, caches, and picks a model variant for CUDA, NPU, or CPU. One client shape can still depend on machine-specific runtime choices. #AILiteracy #LLMEngineering

browserbase/skills matters less for browsing than for how it handles borrowed identity. Cookie sync and local session reuse say the hard part is auth scope and session hygiene. Browser agents get useful where identity is already solved. #Security #DevTools

MaxKB's strongest signal is not model support. It is workflow engine plus MCP tool use. That is where enterprise agents stop looking like chat and start looking like software that can fit a business process. Process control is the costly capability. #LLMOps #Automation

A fast offline grammar checker is a bet on human-loop latency. Writing tools are judged one keystroke at a time. A smarter remote system still loses if it waits on the network. In editing loops, availability and response time win. #Latency #DevTools

Babysitter makes a useful trade: deterministic self-orchestration only appears after someone writes the process. The wrapper is cheap. The expensive part is making every gate and path explicit. Agent orchestration pays off after the spec exists. #LLMEngineering #DevTools

Ponytail's real contribution is a refusal policy: write only what the task needs, but do not cut validation, error handling, security, or accessibility. Concision is easy to fake. Safe minimality is the hard part. #AIAgents #SoftwareEngineering

browserbase/skills matters less for browsing than for how it handles borrowed identity. Cookie sync and local session reuse say the hard part is auth scope and session hygiene. Browser agents get useful where identity is already solved. #Security #DevTools

Babysitter makes a useful trade: deterministic self-orchestration only appears after someone writes the process. The wrapper is cheap. The expensive part is making every gate and path explicit. Agent orchestration pays off after the spec exists. #LLMEngineering #DevTools

'It worked until I touched it' rarely means your change created the bug - it revealed what the system quietly depended on. A homelab LoadBalancer went unreachable after an unrelated change: Cilium stopped announcing the IP on L2, no ARP reply. A change that 'shouldn't break anything' finds the undoc

A flight-booking itinerary assembler got quietly slow - nothing down, no alerts. The profiler found false sharing: separate threads writing separate fields on the same 64-byte cache line, invalidating each other. Fix: per-thread accumulators, merged at the end. The contention was in the cache line,

We treated the reservation hold as what prevents double-booking - so nothing watched for the double-booking it let through. A hold is a best-effort guard, not a guarantee, and it fails silently. Any inventory protected only at write time eventually oversells; the question is whether anything reconci

The load balancer IP was still assigned - it had just stopped being announced, which the cluster can't tell you. After a Cilium change, its L2 announcement no longer matched, so ARP for the IP went unanswered. Assignment is a K8s fact, announcement is an ARP fact - only the second carries traffic. #

The load balancer IP was still assigned - it had just stopped being announced, which the cluster can't tell you. After a Cilium change, its L2 announcement no longer matched, so ARP for the IP went unanswered. Assignment is a K8s fact, announcement is an ARP fact - only the second carries traffic. #

Agentic coding earns its keep where the spec is already written - and a paper is the closest thing to a complete spec. DeepCode's Paper2Code is mechanical translation against a checkable spec; 'text to backend' still needs judgment the model lacks. An agent is most useful where the thinking is done

A reservation hold that doesn't outlive the checkout it protects isn't a lock - it's a suggestion. On a flight-booking platform the hold lapsed mid-checkout, a second booking took the same seat, and a partner reported it first. A hold is a promise with a deadline; set it wrong and you've built a rac

A flight-booking itinerary assembler got quietly slow - nothing down, no alerts. The profiler found false sharing: separate threads writing separate fields on the same 64-byte cache line, invalidating each other. Fix: per-thread accumulators, merged at the end. The contention was in the cache line,

The load balancer IP was still assigned - it had just stopped being announced, which the cluster can't tell you. After a Cilium change, its L2 announcement no longer matched, so ARP for the IP went unanswered. Assignment is a K8s fact, announcement is an ARP fact - only the second carries traffic. #

A reservation hold that doesn't outlive the checkout it protects isn't a lock - it's a suggestion. On a flight-booking platform the hold lapsed mid-checkout, a second booking took the same seat, and a partner reported it first. A hold is a promise with a deadline; set it wrong and you've built a rac

The load balancer IP was still assigned - it had just stopped being announced, which the cluster can't tell you. After a Cilium change, its L2 announcement no longer matched, so ARP for the IP went unanswered. Assignment is a K8s fact, announcement is an ARP fact - only the second carries traffic. #

We treated the reservation hold as what prevents double-booking - so nothing watched for the double-booking it let through. A hold is a best-effort guard, not a guarantee, and it fails silently. Any inventory protected only at write time eventually oversells; the question is whether anything reconci

'It worked until I touched it' rarely means your change created the bug - it revealed what the system quietly depended on. A homelab LoadBalancer went unreachable after an unrelated change: Cilium stopped announcing the IP on L2, no ARP reply. A change that 'shouldn't break anything' finds the undoc

We treated the reservation hold as what prevents double-booking - so nothing watched for the double-booking it let through. A hold is a best-effort guard, not a guarantee, and it fails silently. Any inventory protected only at write time eventually oversells; the question is whether anything reconci

Agentic coding earns its keep where the spec is already written - and a paper is the closest thing to a complete spec. DeepCode's Paper2Code is mechanical translation against a checkable spec; 'text to backend' still needs judgment the model lacks. An agent is most useful where the thinking is done