Zaki Machfj

swe @stripe, prev cs @uwaterloo

[Zaki Machfj @zmachfj]

for folks working in large orgs:

how are you managing https://t.co/qIolRIrB33 (or equivalent) at scale?

specifically curious how teams:
• version common instructions
• prevent drift across tools (Claude, Cursor, Copilot, etc.)
• allow personal overrides without polluting git

[Zaki Machfj @zmachfj]

@xdotli @HamelHusain very nice, will definitely check it out

[Zaki Machfj @zmachfj]

what are the best resources for learning distributed systems and AI infrastructure?

I've read Kleppmann's DDIA: https://t.co/GLc3zgDJvD and implemented some of the systems mentioned in Go like LSM-trees, replica-set with consensus

interested in AI infra/inference

[Zaki Machfj @zmachfj]

summer school term >>>>>
but co-op terms in general were way better

[Demir Eren @demireren_]

co-op term > school term

waterloo people know

[Zaki Machfj @zmachfj]

using AI as a tool to do the job more efficiently, but not dying on the hill

like anything in software engineering, AI is a tool with its use-cases. knowing when and how to use it is a clear edge

[Todd Saunders @toddsaunders]

I have a few friends applying for new jobs and they all ask me the same question.

What exactly does "AI-native" mean.

And as much as I have an answer for them, I really don't.

How would you explain it to them?

[Zaki Machfj @zmachfj]

not reading code for large-scale, critical software is not maintainable imo,

but I can see the argument for startups who are trying to move fast. deploy and fix. can’t say I fully agree but there’s more of an argument there

curious what folks thoughts are here?

[Yahia Bakour @mynameisyahia]

I genuinely do not understand how people ship code without reading it

I use SOTA models for basically everything, max settings, best tools available

They still make hilariously dumb mistakes that would absolutely nuke my infra at any meaningful scale if I just blindly deployed them

Am i missing something?

[Zaki Machfj @zmachfj]

@YessicaAguero agreed, we also had an extensive rubric for judging things like tone, verbosity, etc

and these rubrics can grow and change, but defining them upfront is crucial

[Zaki Machfj @zmachfj]

highly recommend reading this, great example of distributed systems in practice and how to build highly fault-tolerant systems

heavy emphasis on automatic failover, im curious to learn more about semi-sync replication. seems to be crucial for high-availability setups

[PlanetScale @PlanetScale]

PlanetScale's fault tolerance is built on straightforward principles and architectures. The challenge is in the execution.

Here are the principles we follow to keep our systems reliable:

https://t.co/SRC3MiyliM

[Zaki Machfj @zmachfj]

@PlanetScale technical blogs like this for distributed systems are goated, keep em coming 🙏

[Zaki Machfj @zmachfj]

@fardeentwt cursor agent is amazing, and the IDE fallback is necessary for reviewing code easily (ik terminal diff projects exist but not comparable imo)

[Zaki Machfj @zmachfj]

evals are crucial for deploying any LLM or agentic system in production

at a previous job we used mlflow and LLM as a judge for our offline agent evals. for anyone just starting with agent evals these are some great resources

[Xiangyi Li @xdotli]

sharing my personal library on evals 1/n

i put together the highest quality blogs, podcasts, papers, and projects on evals. additions are welcome!

[Zaki Machfj @zmachfj]

@forgebitz not looking at the code doesn't scale for established, large-scale companies with a high bar for latency, reliability and availability

but for startups and early-stage companies who are trying to move fast I can see the argument for it, still not a great idea imo

[Zaki Machfj @zmachfj]

@arushi_ressl wrote evals for enterprise audit agents in a previous role,

llm as a judge and curated datasets are very important

[Zaki Machfj @zmachfj]

@ibocodes ecosystem support and popularity

easier to build, debug and customize when the ecosystem is so much richer for nextjs, although I agree with the premise

[Zaki Machfj @zmachfj]

@sflorimm psychosis

[Zaki Machfj @zmachfj]

@gabriel1 basically /grill-me -> implementation

[Zaki Machfj @zmachfj]

@benhylak say you’re gonna get something done and just do it, on time

very underrated skill

[Zaki Machfj @zmachfj]

@benln codex agent is 🐐

[Zaki Machfj @zmachfj]

@shreypandya might have to try this ngl

[Zaki Machfj @zmachfj]

@rozziitt yea ofc it’s a trade off between token use and quality, depending on whether you’re on a plan or pay per use you may just default to sonnet for everything, but opus on a medium reasoning for planning/research is a reasonable middle ground imo

[Zaki Machfj @zmachfj]

@rozziitt yeah thats basically what im saying

[Zaki Machfj @zmachfj]

@rozziitt opus for coding and sonnet for research? or did you mean the opposite

once the plan is in place, sonnet performs nearly identical for implementation. the bulk of the work is the plan/research phase