Brian Kang

Composer-1 is snarky.. "This plan intentionally keeps auth as SPA-only (your choice). If later you want authenticated Supabase writes (e.g. saved_estimates), we’ll decide whether to use Supabase Auth (or other OIDC provider) or a different backend enforcement model." "Your choice"... dang.

5

Have you heard of SubQuadratic? Yeah me either, until today, and their claims are wild. I was able to land on the waiting list, well see how the claims hold up if I ever get a chance to use it. 12M Token Context window, and claims of 52X faster are bold claims! Time will tell. I will share my experience if I get access. [EDIT], adding this video as its a good break down of the claims. Tim Carambat does a great job breaking it down. https://lnkd.in/gyyJKVMU

1

Next week marks 14 years for me at Buffer, and today we’re launching something I’m genuinely excited about: our new public API. When I first started working on the Buffer iPhone app while freelancing back in 2011 (yeah, it’s been a while), the API was being built alongside the iPhone and Android apps. When it launched, the API was a huge part of the ecosystem. People built integrations, dashboards, automations, experiments, little tools, all the fun stuff that happens when developers can deeply connect with a platform. Over time, the API became partner-only and, honestly, we didn’t give it the love it deserved. Docs drifted, things aged, priorities shifted. So seeing Buffer reopen things feels very full circle. This new API is GraphQL-based, has far better docs, and over time, the goal is to expose much (if not all) of Buffer through it, with new features increasingly being API-first. What excites me most, though, is the timing. We’re in an era where one person with an idea, AI tooling, and a weekend can suddenly build things that previously needed entire teams. Between AI-assisted workflows, automations, custom apps, personal dashboards, and tools we haven’t even thought of yet… it feels like the perfect moment for an open API again. I’ve already been using it heavily myself. For Alpenglow, I've built a Tinder-inspired interface for reviewing user-generated content. Photos land in Buffer drafts, and I can quickly swipe through them, moving the best ones into the queue to keep content flowing. I’ve also been experimenting with eInk displays and built a Buffer integration for TRMNL that shows queue counts, upcoming posts, and recently sent posts right on my desk. I've also added Buffer as an OAuth provider in the TRMNL plugin system so people can quickly build their own plugins and displays on top of it. I also pulled together the landing page for the API launch. We had some fun with the background effects, although I suspect most engineers will click straight through to the docs and immediately start tinkering 😅 It was also built without touching a single line of code. Thanks, Claude! Excited to see what everyone builds with it 🚀 https://buffer.com/api #BufferAPI

54

2 Comments

🐶 Ever wonder what your dog is actually thinking? I decided to build and train an AI LLM to find out. Meet Rosetta Bone 🦴— a custom, fine-tuned LLM designed to write from a dog's first-person, sensory point of view! While you can prompt a massive frontier model to "act like a dog," you usually just get a sophisticated AI performing dog-like dumbness. As the context grows, the persona drops, and it reverts back to a human's visual perspective rather than a dog's scent-and-sound-driven reality. I wanted something more authentic, so I built and trained my own. Here is a look under the hood of how I did it: 🔬 The Training Data: To get the perspective just right, I curated a custom training dataset blending three unique sources. The model is grounded in dog-POV fiction (like Call of the Wild and Beautiful Joe), actual canine olfaction research papers, and real dog-behavior Q&As. 🧠 The Tech Stack: It's built on top of Meta-Llama-3.1-8B-Instruct (4-bit) and I fine-tuned it using mlx-lm LoRA. 💻 100% Local: No API inference bills here! By leveraging Apple Silicon, I trained the adapter entirely locally on my Mac, meaning I can run my dog-brained AI forever for free. The result is a model where the cadence and sensory details are baked directly into the weights. It is excitable, easily distracted, and confidently wrong about cause-and-effect—exactly like a Very Good Boy. You can check out the training pipeline, the code, and some hilarious sample outputs (like how it reacts to "the smell of bacon") over on GitHub: 👉 https://lnkd.in/g9u-XcyA What everyday stimulus should I ask Rosetta Bone to react to next? Drop your suggestions in the comments! 👇 #MachineLearning #LLMs #ArtificialIntelligence #AppleSilicon #FineTuning #OpenSource #Llama3 #GenerativeAI #Python #RosettaBone

14

1 Comment

I’ve been spending quite a bit of time working with Claude Code lately, and ran into a recurring limitation: a single global configuration for everything. Different contexts (work, personal projects, experiments) often require different instructions, tools, and behaviors — but there’s no clean way to isolate or switch between them. That led me to build CVM (Claude Version Manager). The idea is simple: bring a model similar to nvm (Node Version Manager) or gvm (Go Version Manager) to Claude Code environments. With CVM, you can: - Define multiple, fully isolated Claude Code profiles - Switch between them with a single command - Scope them globally or per project - Version and sync them via Git - Persist knowledge across sessions in a structured way It’s intentionally minimal — no abstractions beyond what’s needed to manage and switch contexts reliably. I’ve been using it with a personal profile, but the more interesting direction is treating profiles as shareable artifacts. If you’ve been experimenting with structured setups around Claude Code, I’d be interested in seeing how others approach this — and whether sharing reusable profiles makes sense as a pattern. Repository: github.com/chichex/cvm

18

TOO EARLY TO CALL From friend: ---- To compare and learn, I built an app in parallel on both Codex and Claude Code. - Claude Code was far superior in design. - But Codex beat Claude Code by orders of magnitude when it came to development and (in particular) testing. I had to spend a fraction of time on testing the app in Codex vs Claude. Codex was far better at identifying and fixing bugs before it came to me for testing. Claude had many more bugs, and its testing was limited and lower fidelity. ---- tl;dr It's too early to pronounce a singular winner in AI software development (let alone in other more nascent spaces). We have multiple excellent companies going neck-to-neck with zero complacency and incredible urgency. Developers will be the ultimate winner as Anthropic and OpenAI (and Google and Factory) continually one-up each other across different dimensions.

214

39 Comments

From Code to Capital 💼💡 Moving into a Principal Engineer role has been a big shift—from writing code every day to now owning technical direction, governance, and mentoring other developers. But one thing became clear: - To design better systems in FinTech, I need to deeply understand the financial world they serve. That’s why I’ve started preparing for the CFA—yes, the Chartered Financial Analyst path. Not because I want to become a portfolio manager, but because: - 🔍 I want to know how finance really works, not just the APIs that power it - 💸 I want to understand the buy side vs sell side, investment strategies, and what drives the products we help build - 📈 I want to bridge the gap between tech and finance, and speak both languages fluently Studying finance has already started changing how I think about product value, data, and risk. It's helping me grow not just as an engineer, but as a tech leader who gets the business. If you’re in tech and working in or near finance, consider learning the fundamentals. It’s humbling—and powerful. #PrincipalEngineer #FinTech #CFAJourney #EngineeringLeadership #BuySide #SellSide #FinanceAndTech #CrossDisciplinaryLearning

43

3 Comments

I built a Vampire Survivors clone to test Cheng Lou's 'Pretext' library. That sentence should tell you everything about how deep I went with it. The thing that got me: AI-generated UIs need deterministic layout. You can't have an agent building interfaces that reflow unpredictably. Pretext does layout in pure arithmetic, zero DOM reads. That's not a performance trick. That's a different contract with the browser entirely. You can try the game here if you like: https://lnkd.in/e-mUdX8P

15

4 Comments

A few things shipped for tokf this week. The headline: a community filter catalog at tokf.net/filters. If someone has already figured out the best way to compress `kubectl get pods` output for an LLM, you shouldn't have to reinvent it. You can find filters two ways — from the CLI: tokf search kubectl tokf install <hash> or browse the web catalog by command pattern, savings %, or usage count. Each filter page shows the full TOML with syntax highlighting so you can read it before installing. Lua filters are sandboxed and inlined at publish time — what you see is what runs. tokf also now works with OpenCode and OpenAI Codex CLI, not just Claude Code. The hook installer handles all three. A few other things that landed: — make/just support via shell-override wrappers (the model runs make targets too) — env var prefix stripping, so NODE_ENV=test npm test matches the npm filter correctly — account deletion: tokf auth delete-account removes your data immediately. Published filters stay, anonymized. — Terms of Service: tokf.net/terms tokf.net/filters

13

Vibe Programming For Real: Part 1 I built a complete medical records app in 3 weeks. Zero lines of code written manually. Cost: $0. Here's what I learned about "vibe programming" with AI: --- The project: I had a specialist appointment for a chronic cough I've had for 2+ years. I wanted to bring complete medical records—organized, analyzed, and ready for discussion. So I built an app: Cloudflare backend, 850+ clinical concepts extracted from 34 documents, specialist briefing with timeline and differential diagnosis. All code generated by Windsurf/Cascade (Claude Sonnet). I just told it what to build. --- What worked: 1️⃣ Stop reading AI-generated code The code was correct. Reading it line-by-line added no value. My job became: verify outcomes, not review syntax. 2️⃣ PROJECT_PLAN.md is your steering wheel I broke work into phases → steps → tasks. Never proceeded without verification. Marked complete only after testing with real data. 3️⃣ Journaling everything Every prompt, every decision, every problem—logged. This became my audit trail and the basis for this post. 4️⃣ Git commits after every step Vibe programming moves fast. Sometimes too fast. Small commits = easy rollbacks. --- What didn't work: ❌ Model switching mid-project Tried Gemini 3.0 (ran off adding features I didn't ask for), AntiGravity (corrupted my files), free Grok (copied irrelevant code). Came back to Claude. ❌ Expecting complete output from incomplete data The specialist briefing was only as good as my records. Turns out my doctor didn't document most of my symptom discussions. --- Key lessons: ✅ Vibe programming works—but requires expertise I still did requirements engineering, architecture decisions, verification, debugging. I just used English instead of code. ✅ Workflow and Tools The tools mattered more than I thought. But it's the verifiable steps, never skipping verification, journaling, and Git safety net that made it work. ✅ AI amplifies your inputs Missing data = missing output. AI doesn't fill gaps—it exposes them. --- What's next: Getting insurance claims data to find all my visits, then pulling records from every provider. Stay tuned for Part 2: "Separation of Concerns: Empowered by an Intelligent Code Machine" --- What's your experience with AI-assisted development? I'd love to hear what workflows are working for you. #VibeProgramming #AI #SoftwareEngineering #Healthcare #BuildInPublic Full article: https://lnkd.in/gRQZnB-D Code: https://lnkd.in/gDvFm-my

17

1 Comment

I find Copilot a reasonably helpful tool, specially when we have to write repetitive stances, like test cases, but it visibly lacks understanding of even the most simple types in TypeScript (which is unbelievable, since it also belongs to Micro$oft, that owns GitHub): it suggests unexpected properties, often wrongly guess the value types, and so on. This article gives a much deeper look on what LLMs are really capable of. https://lnkd.in/dymfd5da

6

I started a newsletter about AI-assisted engineering. Not hype. Not "10x your productivity with this one prompt." Just structured approaches that actually work in real codebases. Planning before execution. Types as constraints. Tests as verification. For engineers who want AI speed without losing control. It's called Controlled Acceleration. https://lnkd.in/du-NBDPY

34

5 Comments

v1.1.0 is out for the Mobile Observability Plugin for Claude Code. Flutter support just shipped, along with three new vendor integrations. New vendor guides: → Firebase (Crashlytics, Analytics, Performance Monitoring, Remote Config) → OpenTelemetry (opentelemetry-swift v2.3.0, opentelemetry-android v1.0.0-rc.1) → Measure.sh (open source and self-hosted, for teams that need data sovereignty) New anti-pattern detection: localized strings in telemetry. Your crash groups shouldn't split by locale. "Erreur réseau" and "Network error" are the same bug. The plugin now catches this in Swift, Kotlin, TypeScript, and Dart before you ship it. What this plugin does: Run /instrument ios or /instrument flutter and Claude analyzes your codebase, finds the gaps, and generates a plan. No more guessing what to measure. Install it: claude plugin add calube/mobile-observability I also put together a demo video walking through how it works: https://lnkd.in/gheM77Zb Shoutout to James Sedlacek for open sourcing his Scrumdinger application 👏 Huge thank you for the feedback and support. It's been awesome to see 🫶

62

7 Comments

We've come a long way from "the AI can write tests for you" of just a year ago! The vast majority of our code is written with AI, the day-to-day for software engineers is totally different, and we have coding agents working around-the-clock in the cloud (even more on that soon). Very proud of what, and especially how, we have accomplish all of that at Plaid.

29

I wanted to see how close I could get Claude to ~one-shot a typical startup while allowing for high-signal user input. I ended up with two Claude Code skills that handle the heavy lifting: Go live checklist — 9 parallel agents scan your repo for billing, auth, security, testing, analytics, error handling, and more. It compiles findings, works through decisions with you interactively, and outputs a prioritized checklist. Production audit — picks up that checklist and enters an autonomous loop: discover workflows → assess (UX, security, performance) → split into conflict-free work packages → fix everything in parallel via git worktrees → re-assess → repeat until clean. The whole thing is checkpoint-based, so it survives context flushes and session restarts. And it works across web, iOS, Android, React Native, and Flutter Find it at https://lnkd.in/gbwz2yZ2. Contributions welcome!

33

1 Comment

I just released Ace Template, which allows you to go from `rails new` to full-featured deployed Rails app in under 5 minutes! https://lnkd.in/g3YYUqxb Demos 👇 - Kamal https://lnkd.in/gJ_kcRAk - Fly https://lnkd.in/g3PHkctz - Heroku https://lnkd.in/gn-qXQYk Ace Template will configure Postgres/MySQL/SQLite for you, as well as Redis if you so choose. Many more options are available. The template is designed to keep up with future versions of Rails by running daily tests against Rails `main`.

7