Thanks! I'm glad someone appreciated it, I spent a bunch of time getting something that looked nice and also tuning the UX e.g. for highlighting the marker when you hover the footnote and vice versa!
I only hit the weekly rate limit on once, one day before it was going to refresh anyway. Apart from that, never hit another limit.
AI has no sense of time. It sees a codebase as a snapshot but doesn't feel why decisions were made, reversed, or avoided. Losing a senior engineer hurts a team for similar reasons; AI never had that institutional memory to begin with.
At several points, I lost the mental model of the codebase. I had become the engineering manager who doesn't actually understand the code, the exact person I used to complain about. AI makes it dangerous to stop owning the "why" behind your implementation.
AI coding is a slot machine. You prompt, wait, and hope for a win. At 1am, I’d think "one more prompt" instead of actually solving the problem. When I got tired, my prompts got vague and the output got worse, but I’d just pull the lever again anyway.
I just published a deep-dive into the 250-hour build behind syntaqlite, a SQLite formatter and LSP I built using AI agents.
AI agents were the only reason built this after 8 years of wanting but there's a psychological toll to AI-assisted engineering.
The post-mortem:
lalitm.com/post/buildin...
The result is a CLI and LSP that gives you rustc-style diagnostics for things like CTE column mismatches and function typos, all without needing an active database connection.
I wrote a full breakdown of the implementation and why 1:1 fidelity matters so much to me here:
lalitm.com/post/syntaql...
Writing a hand-rolled parser is where so many tools fall over, so syntaqlite uses SQLite’s own Lemon-generated grammar, adapted into a C/Rust stack.
If SQLite accepts it, syntaqlite does too. I tested it against 396k statements from the official SQLite test suite and got 99.7% agreement.
For a long time, I’ve wanted a reliable, fast formatter and LSP for SQLite SQL. One that actually models SQLite’s weird/specific features, like virtual tables, application-defined functions, and the various flags that change the language.
So I started building syntaqlite.
Hit a case in Perfetto where even the venerable binary search was still too slow! I ended up discovering an algorithm that I hadn't heard of before called exponential search. Wrote up what I learned at lalitm.com/post/exponen...
They basically solve two problems caused by AI coding agents:
1) Sandboxing to reduce the chances prompt injection ruining your life
2) Having "clean slates" where you can prototype freely and fast
All in all, I'm semi-won over. Won't be paying every month but it's a tool to add to the toolbelt.
Initially, I was very confused why these were getting so much attention given that Linux VPSes have been a thing for 2 decades but after trying them out and reflecting, I think I have an answer.
Just wrote up some thoughts after trying out all the "VPS as a service" products which reached the Hacker News frontpage over the past month - lalitm.com/trying-sprit...
venn diagram. circles labeled "Angled lines" and "Different angled lines", they are filled with patterns of slightly different angled lines. the intersecting area is labeled Moire pattern and indeed has a moire pattern in it.
I made a venn diagram to help you understand Moiré patterns.
It's the guide I wish I had when I first started 3 years ago.
It covers how to build a comprehensive plain-text accounting system with Beancount. We're talking 18 accounts, 3 currencies and maximal automation.
🔗 lalitm.com/post/one-num...
Been quiet for the last few weeks because I was cooking this up!
I know it's resolution season, so if you're engineering-minded and want absolute control over your money (handling everything from daily spend to cross-border investments) I wrote a blog post for you! 👇
You don't have to be a "Solver" dropping into fires. You can be an "Architect" building the foundation.
It’s okay to trade external validation for deep technical ownership.
Full essay here: lalitm.com/software-eng...
The mental model that saves me is the "Shadow Hierarchy."
I don't worry about impressing my VP. I worry about the Staff engineers in Pixel and Chrome telling their VPs: "We literally cannot debug the next phone without this tool."
That is un-fakeable influence.
In product, speed is currency. In infra, context is currency.
If you rotate teams every 18 months to chase the "spotlight," you never gain the pattern matching required to solve systemic problems. You fix bugs, but you don't build leverage.
I’ve been reading a lot of Staff+ advice that says you need to be "fungible," chase executive priorities, and move fast.
As a Senior Staff engineer in infra/devtools, I found that advice is a recipe for burnout.
The secret sauce: gamification (leaderboard + t-shirts), hard limits (no bug >2 days), and critical mass (~40 people creates real momentum).
#EngineeringManagement #SoftwareEngineering #ProductProductivity
Burndown graph of bugs fixed
The results? 189 bugs fixed. A 2021 feature request finally implemented (took just 1 day!). Team morale through the roof. And users actually notice the polish.
Every quarter, our 45-person eng team stops ALL roadmap work for an entire week. No new features. No meetings. Just fixing the small stuff that's been annoying us and our users.
I'm a bit ashamed to say it's only today I learned what the number in Linux man page titles (e.g. the "3" in `sleep(3)`) is: lalitm.com/til-number-i...
Just sent my latest newsletter rounding up the last week of posts to my blog: lalitm.substack.com/p/when-good-...
After reading a fascinating article called "Don't Build An Audience", I've been doing a lot of soul searching about the type of technical writing I like and how to get people to see it. Wrote down my thoughts in a new post: lalitm.com/on-why-i-wri...
Just sent my latest newsletter rounding up the last week of posts to my blog: lalitm.substack.com/p/castx-as-s...