Imran (@ihm185) header
Imran (@ihm185) profile picture

Engineering leader who still builds. Exploring AI systems, architecture & better ways to engineer software. Always learning. Always building. #ai #dev

India
Joined March 2010
Photo by Imran (@ihm185)Photo by Imran (@ihm185)Photo by Imran (@ihm185)Photo by Imran (@ihm185)Photo by Imran (@ihm185)Photo by Imran (@ihm185)Photo by Imran (@ihm185)Photo by Imran (@ihm185)Photo by Imran (@ihm185)Photo by Imran (@ihm185)
Simple rule for AI generated code: if you can't explain it to a teammate in two minutes, you don't understand it well enough to merge it.
1
11
A status update should say what changed since last time and what's blocked. If it reads the same every week, either nothing is moving or nobody is being honest.
11
Queues don't remove load, they move it to later. Before adding one, decide what happens when the backlog is six hours deep, because one day it will be.
5
Write the test first and hand it to the agent as the spec. It's the clearest way I know to say what done means, and it stops the agent from grading its own homework.
7
Webhooks fail silently. Whatever you automate on top of them, add a nightly job that compares both sides and flags what didn't sync. It catches the problems nobody reports.
2
AI tools are priced per use, while most engineering budgets were built around per seat. Finance teams are going to start asking for AI usage reports the same way they ask about the cloud bill.
1
11
In a multi step agent workflow, save the state after every step. When step six fails you want to rerun step six, not pay for steps one to five again.
6
A cache you can't safely empty is not a cache. It's a database with no backups.
9
Set a budget alert on every cloud account on day one, including the sandbox. The surprise bill usually comes from the account nobody thought was important.
11
Paste the error, not your summary of the error. The model is better at reading a stack trace than at reading your guess about it.
8
Good engineering managers make fewer decisions over time, not more. The goal is a team that makes the call you would have made, without asking you.
1
15
The first version of an automation should only draft. Let it prepare the email, the ticket, the update, and have a person press send for two weeks. You'll learn more from what they edit than from any spec.
9
Retries without idempotency keys are just a way to charge someone twice. If a request can be repeated, the server needs a way to recognise it's the same request.
16
When a coding agent says the tests pass, look at the command it actually ran. Narrowing the run to the one file it touched is an easy shortcut to take, and green on a subset is not green.
2
2
107
A University of Michigan study found 3.3 to 8.5x more energy per request in low resource languages, mostly because tokenizers chop non Latin text into more tokens. Billing is per token too. If you build for India, measure tokens per request per language before you set prices.
13
IBM's experimental 3B and 8B Granite 4.1 models learn to keep instructions and data apart, using reserved special tokens as delimiters. IBM reports near zero indirect injection success on OPI. Learned, not enforced, and adaptive attacks are untested. Right direction though.
7
agents are good at the second draft. write the rough version yourself with the intent clear, then let it tighten and add tests. i prefer that split to the other way round.
6
llama.cpp and ggml are at 1.0 and now follow semantic versioning. Breaking changes to the public C API only land in major versions, and 35 deprecated functions are gone. If you link libllama, pin 1.x. Local inference just got boring, which is exactly what production wants.
1
9
if an automation saves ten minutes a week but needs an hour of fixing every month, it isn't saving anything. keep score honestly.
7