Sachin Nene

"I have only made this letter longer because I have not had the time to make it shorter." - Blaise Pascal'' This also applies to software demos. We've tightened the onboarding workflow and now have a 60 second demo of the journey from a blank slate in Claude Code to a deployed preview application on VeilStream.

2

1 Comment

To date, writing extensions for Envoy has been a laborious process that involves writing C++ and compiling a fully custom build of the entire proxy.  Can we make that easy and make Envoy more approachable by developers?. Can we create an ecosystem of reusable extensions around it - akin to how npm ecosystem fundamentally altered Node.js's trajectory. With that vision, we at Tetrate have been working with Envoy community to first make Envoy easy to extend by any developer. We added dynamic modules a year ago and then added a proper Go and Rust SDK around it. Today, we are making the developer experience much easier and announcing a home for Envoy extensions with an open source extension marketplace around it. Thanks Matt Klein Anna Berenberg for your support and thanks Rohit Agrawal Ben Plotnick Anubhav Aeron for early feedback and trials. You can read more about Built on Envoy here: https://lnkd.in/gbNz9-3v and see a cool demo video from our engineering lead Ignasi Barrera on how to write a simple Go extension and test it locally for Envoy to not allow AI calls to specific models. https://lnkd.in/gcpYJyjD

74

I thought this was a good and very reasonable talk by Eleanor Millman on how to prioritize projects for a platform engineering team: https://lnkd.in/gaYtUHRH A few insights I took from it: - Unlike product features which can be more directly tied to revenue, platform engineering work is a couple steps removed. But that doesn't mean it's unimportant! Far from it, platform engineering projects can make everyone at the company more efficient, able to produce higher quality software, etc. - While "urgency" is always going to be a factor, you don't want to just be putting out fires, you want be able to prioritize work which is "high impact" ... even if that impact isn't felt immediately. - Good rule of thumb: prioritize "highest impact for lowest effort". - The talk has some ideas on what factors to choose for impact, and how to blend them. For instance "speed of development" is one factor, "cloud cost optimization" might be another. The weighting of different impact factors can change over time, depending on the needs of the business. I would say that a lot of companies don't have much methodology here but I can really see the value in codifying it. You can always change the methodology or the weighting if it's spitting out results that don't pass the smell test or you really feel it is leading the company astray. Clarity can give the org more freedom to put resources behind projects that would otherwise never be taken on. When things are unclear, prioritization still happens implicitly, but the decisions tend to be a lot more random and fear-based and the org is worse off as a result.

15

3 Comments

MSK (Managed Kafka) looks simple on the bill until you pull it apart. Most teams know they pay for broker hours, which range from ~$0.043/hr for a kafka.t3.small to over $10/hr for a kafka.m5.24xlarge. A common production setup with three kafka.m5.large brokers costs around $460/month — that's 3 brokers × $0.21/hr × 730 hours — just for compute. However, many overlook the ongoing costs of storage at $0.10 per GB per month and the potential charges for data transfer, which is $0.01 per GB each way for cross-Availability Zone (AZ) traffic. An often missed aspect is the waste incurred from idle clusters. If no messages are processed over a seven-day period, as detected by CloudWatch, the cluster is effectively idle — and unlike EC2, MSK clusters cannot be stopped, only deleted. Common scenarios that lead to idle clusters include deprecated streaming pipelines or development clusters mirroring production setups left running without a purpose. Taking a moment to regularly review cluster activity can prevent unnecessary charges and ensure that resources align with current application needs. CloudWise detects idle MSK clusters automatically so you don't have to audit them manually. #AWS #AWSMKS #CloudWise #CostOptimization #FinOps

5

1 Comment

If you’re in/near Atlanta, don’t miss the Agentic AI meetup by AICamp this Wednesday, Sep 18 at Improving (Alpharetta, GA). If you’re building with GenAI/LLMs or exploring agent patterns, expect deep-dive talks, hands-on code labs, and plenty of hallway learning. 💻 My colleague Michael Hall, Principal Developer Evangelist, will present “Device to Datacenter: How Mobile & Cloud AI Coexist on Arm.” He’ll cover when to run inference on-device vs. in the cloud (or hybrid), and how to design for portability and performance across the stack. https://okt.to/UbRIA4

8

Six months ago, Spotnana engineering committed fully to Claude Code and Cursor. Adoption happened quickly and naturally. What surprised us most was a feature released last July in that many developers still haven't discovered: custom sub-agents. When you use ClaudeCode in its default form, you're working with a capable generalist. Custom sub-agents let you create specialists instead. You define them by placing a markdown file in the .claude/agents/ directory with simple YAML frontmatter that sets the agent's name, persona, and allowed tools. Claude then automatically delegates tasks to the appropriate specialist based on context and no manual prompting required. Each sub-agent runs in its own isolated context window, so your main conversation stays clean. The agents are also shareable across projects, which lets institutional knowledge build over time. The biggest challenge we faced was getting routing reliable. When agent descriptions overlap even slightly, Claude struggles to choose the right one. We've learned to make each persona description extremely precise and distinct. Another lesson came the hard way: sub-agents inherit your permission settings by default. I once gave a research-focused agent write access, and it began refactoring code while trying to investigate a bug. One improvement I would love to see is a clear post-session audit log that shows which agent handled which actions and when. Right now, tracing responsibility requires digging through long transcripts, which slows down debugging and makes accountability harder. If you're already using Claude Code (or Cursor with Claude) and haven't tried the /agents feature yet, you're missing out on significant capability. What sub-agents has your team built, and how have they changed your workflow? #ClaudeCode #AICoding #AgenticAI #CursorAI #DeveloperTools

59

3 Comments

The ROI isn’t just speed—it’s transfer. Prompting patterns I used for: - Survey design → also work for interview guides. - Terraform reviews → also work for policy checks. - Meeting digests → also work for PRDs. Learn once, apply everywhere.

5

In this presentation, Colin cogently articulates the rationale for choosing Components over Containers and why Agentic Systems provide a catalyst for that paradigm shift (especially from around the 19 minute mark).

2

We're excited to announce that all the ADBC drivers Columnar distributes can now be used in Java. The latest ADBC release adds a JNI layer that lets Java applications load drivers, run queries, and efficiently fetch results in Apache Arrow format. David Li and I wrote a blog post with quickstart examples and a performance comparison to JDBC. Link in comments 👇

163

15 Comments

From Claude Code for developers to Claude Cowork for everyone else—this is a real shift. Stop controlling prompts and start controlling outcomes. Instead of just chatting, you give Claude access to a local folder so it can read, edit, and create files to tackle multi-step tasks like organizing docs, building spreadsheets, or drafting reports. It’s the move from “assistant” to true AI agents at work. Sure, it has bugs—but directionally, it’s very promising. #AI #Claude #vibecoding https://lnkd.in/ghmFtTFp

6

2 Comments

Claude Code and Opus 4.5 have recently made meaningful strides in raising the bar on agentic development. However, I’ve discovered that when you treat the agents more as teammates and less as automatons they really excel. Let me explain. Claude Code’s GitHub integration finally lets me manage agents the way a great CTO manages developers: with clear objectives, structured workflows, nuanced guardrails, and most importantly, stable context. I start every project in Claude Code's planning mode now. The agents and I work together to draft comprehensive plans. I review. We iterate. Then, and this is the key, we instantiate those plans as GitHub epics. Each epic becomes a container for related issues. Each issue becomes a discrete task with clear objectives, architectural guidelines, and acceptance criteria. The agents work methodically through the issues and submit PRs when done. They even check each other’s work, just like a human team would. Because Claude Code maintains context through these epics, we’re no longer fighting context window exhaustion. A three-week project stays coherent from day one to deployment. The agents can see the overall arc of the project at any time, track progress, and recall the architectural decisions from week one as they implement features in week three. The results have been striking. Code quality has improved markedly. The agents are producing more consistent, better-structured code that actually follows our established patterns. Error rates have dropped to levels I’d expect from more senior developers, not the interns agents sometimes seem to channel. And if they stray, I just ask them to revisit the epic and validate whether they’ve delivered against the requirements. They usually course-correct. With a growing level of comfort, I’m now starting to let the agents monitor GitHub on their own and select their own epics to work on, updating the underlying GitHub issues as they make progress. They’re not just executing rote tasks anymore—they’re participating in overall project management. Great tech leaders don’t micromanage. They set clear objectives, establish workflows, remove blockers, and trust their teams to deliver. That’s precisely what I’m doing with my agent teammates now. I’m just applying decades of engineering management wisdom to a new kind of hybrid team. And it’s working. Carpe Agentem. 🎯 #AgenticDevelopment #GitHubIntegration #ClaudeCode #EngineeringManagement #AI #FutureOfWork

168

15 Comments

We are going all in on agentic coding. But this requires tight controls, right infrastructure, shared internal knowledge, and verification processes. I wrote about our current process - the planning framework, how we handle parallel execution, code review loops, and the lessons learned along the way. Read the full blog post here: https://lnkd.in/gwXSFXwu

38

1 Comment

LLM state management for a text rpg (...not usually my jam, but I want to use it to help me learn classical languages better) and LLM state management for a personal assistant have a lot of overlaps, and working on them in parallel has repeatedly helped me make breakthroughs on the other side. JSON documents, as a whole - not specific fields, like Mongo Vector search - should be trivially handled by embeddings and vector search. Not quite 'done', but here's some progress.

2

Really proud of the work Harry Pierson and the team have done in releasing DBOS Java today! DBOS Java brings lightweight durable workflows to Java. Workflows make your program durable by checkpointing its state in Postgres. If your program ever fails, when it restarts all your workflows will automatically resume from the last completed step. Workflows are useful for solving many problems, such as: - Orchestrating business processes so they seamlessly recover from any failure - Operating an AI agent, or anything that connects to an unreliable or non-deterministic API. - Running reliable background jobs with no timeouts. - Processing incoming events (e.g. from Kafka) exactly once What’s unique about DBOS workflows is that they’re implemented entirely in a lightweight Postgres-backed library. All you have to do to use DBOS is install the library into your application and register workflows and steps. As your program runs, the library checkpoints your workflows and steps in Postgres. If your program ever fails, when it restarts it uses those checkpoints to resume all your workflows from the last completed step. There are no other dependencies you have to manage, no separate workflow orchestrator–just your program and Postgres. The big advantage of this approach is that you can incrementally add DBOS to any Java application–it’s just a library. You can add workflows to your application without rearchitecting it or adopting any new infrastructure. It works out of the box with frameworks like Spring. Also, because it’s all Postgres-backed, you get all the tooling you’re familiar with: backups, GUIs, CLI tools–it all just works. You can use Supabase, RDS, Neon, or any other Postgres provider. Check it out on GitHub, try it out, and give us a star!

49

6 Comments

If you’re a CEO/founder or PE operator and your tech org is: ❌ Shipping late ❌ Stuck in reactive mode ❌ Struggling through the Series A/B jump, or post-acquisition integration or after a major release It’s not “a delivery problem.” It’s an operating model problem. At the inflection point, the old playbook breaks: what worked at 10 engineers fails at 40… and breaks again across multiple products/teams. That’s where I come in as a fractional CTO: - Spot platform and delivery risks before they hit revenue and retention - Rebuild execution rituals (prioritization, ownership, release confidence) without derailing the roadmap - Align architecture + team structure to your value-creation plan (growth, margin, or exit timeline) If you’re feeling the strain of growth, I can help you turn “busy and reactive” into predictable shipping, and a platform that scales. #CTO #SaaS #StartupScale #PrivateEquity #TechStrategy

8

3 Comments