Raphael Santo Domingo

Spent some time digging into why prompt caching breaks in agent loops. Subtle prompt serialization changes (especially around tool definitions) can drop cache hit rate to zero. TL;DR: System + tool schemas are a contract. Late bind IDs/metadata. Full writeup and experiments: https://lnkd.in/e_Qngs5y

11

How are folks solving the rapidly evolving situation of non-tech or semi-tech folks using things like ChatGPT to automate their toil and doing so in ways that are to at least some extent anti-patterns? So I'm asking: do you make these folks commit to github? do you run static analysis on that code? do you run something like endpoint security management with special features now turned on? do you have an overview committee? etc etc. #cantcontrolpeoplereally

4

3 Comments

Mault will be releasing our multi-agent flow this week. It includes: Does the code work? **Agents operate in scoped sandboxes (worktrees/PIDs) so they never bump into each other. ** Gatekeeper protocol mandates Hexagonal Architecture, SRP, and zero hard-coding. ** Pre-commit hooks enforce testing structure and code quality before it leaves machine. ** CI serves as the ultimate arbiter, utilizing mutation testing and stealth security checks to ensure 'correctness' isn't just a pass/fail metric. The Results: --> 2.6% Code Duplication (industry average is often 10-20% in rapid-growth phases). --> 40+ PRs per day with total architectural consistency. Does the product work? ** Now that code is proven to work, we have a tester agent tasked with proving the product, and user journey to work (will demo that seperately)

11

We weren't solving the right problem in code reviews. Our engineers were spending hours catching accessibility violations, security patterns, and architectural anti-patterns during PR reviews. Not debating design decisions. Not mentoring. Just... repetitive pattern detection that machines should handle. The real cost? Senior engineers burning cognitive capacity on checklist items instead of strategic thinking. So we built something different: an agentic code review system that learns from every merge request outcome. What makes this agentic (not just "AI-assisted"): 🤖 7+ language-specific agents with specialized pattern detection - each agent understands the idioms and anti-patterns of TypeScript, Java, Python 🧠 Autonomous learning pipeline - extracts patterns from review history, validates them, and continuously improves from every interaction 🎯 Auto-approval agent with confidence thresholds - escalates to humans only when needed (≥0.7 confidence, ≥0.8 accuracy, ≥5 observations) 📊 Self-correcting through feedback - yes, there are false positives initially, but that's the point: every correction teaches the system, reducing noise over time The real shift: moving feedback left We're now building an IDE plugin that catches these patterns during development, not during review. Why wait days for a PR comment when you can get feedback in seconds while you're still in flow state? This is the compounding effect of agentic systems: faster feedback loops → more learning cycles → better pattern detection → less cognitive load on engineers. The transformation: Engineers now orchestrate quality rather than enforce it manually. They set standards, validate edge cases, and focus on architecture - while agents handle pattern detection at scale and learn from their corrections. This is what agentic SDLC looks like in practice: not replacing engineers, but fundamentally changing how they allocate their time and expertise. The question every engineering leader should be asking: Are you building productivity tools, or are you transforming how your engineers think?

17

10 Comments

50 MB/s ingest, zero-copy reads for tailing consumers. Then a backfill request for 7 days of history hits. Suddenly, your p99 produce latency spikes from 2ms to 250ms. 📉 This isn't a network issue. It's the "Catch-Up Tax" – a memory management collision within the Linux Kernel. Your cluster is effectively DDoS-ing itself. What we're seeing at Azguards Technolabs is that Kafka, by design, outsources its read cache to the OS Page Cache. When cold, historical data floods this cache during a backfill, it ruthlessly evicts the hot, active segments your real-time consumers desperately need. Your critical data paths are forced onto physical disk. The definitive solution isn't just kernel tuning or bandwidth quotas. It's an architectural shift: Tiered Storage (Kafka 3.6+). ⚙️ By offloading historical segments to object storage like S3, your local NVMe disk transforms into a high-performance "Short-Term Cache" for active data. This physically isolates conflicting I/O patterns, maintaining predictable p99 latency even under a "thundering herd" historical read. Reliability isn't just about raw throughput; it's about predictable performance under stress. We re-architect Kafka environments to decouple these conflicting access patterns, ensuring your critical data flows remain unaffected. 🚀 Is your Kafka infrastructure buckling under historical reads? Read the full technical breakdown here: https://lnkd.in/gsZ7sJKE #Kafka #ApacheKafka #Infrastructure #DataEngineering #CloudNative #Performance #TieredStorage

2

👨‍💻 We don’t micromanage at Optavyn. Every dev here knows they’re not “just writing code”, they’re building products that matter. Every feature we ship solves a real-world problem. Every API we build connects with a user’s workflow. Every screen we design has intent behind it. Our team works remote, but we’re aligned, focused, and committed to the end goal: Delivering results not just tasks. If you're a founder or agency who values clear ownership, clean code, and reliable delivery, we might be the team you’ve been looking for. #TeamCulture #RemoteTeam #SoftwareProduct #OptavynCulture #ExecutionMatters #TechThatWorks

4