Ajmal Aksar

Sr. AI engineer @ Express Analytics. I test AI agents on real work and post the receipts, including when they lose. Building kAInet, ads that wait for yoυr ΟΚ.

[Ajmal Aksar @AjmalAksar]

been building kAInet for a little over a year. pilot is open, no card needed. what's the one weekly ad task you'd most like off your plate? tell me, or dm me for access or a walkthrough

[kAInet @kAInet_tech]

kAInet builds your Google and Meta ad campaigns. Built by AI. Launched by you.

Here's how it works ↓

[Ajmal Aksar @AjmalAksar]

@hwchase17 would love to try it out

[Ajmal Aksar @AjmalAksar]

@omarsar0 it always comes back to rag. in my experience naive rag plus a newer model works really well most of the time. i think classical rag was partly held back by the model, not the idea. the retrieval was fine, the reader just wasn't good enough

[Ajmal Aksar @AjmalAksar]

@stellakaval the win for me is parallel work in very little time, with one main orchestrator still steering. subagents chase the tangents, the main layer keeps the autonomy and redirects when needed

[Ajmal Aksar @AjmalAksar]

@cognition @devindevelopers @Avi31m @hamzehaq7

do I win? 🫣

https://t.co/ESwaS0wC20

[Ajmal Aksar @AjmalAksar]

newer image models make getting a creative easy. getting one that feels lifelike takes engineering imperfection: knowing what makes an image feel human, then prompting so agents create the imperfection the natural world has. sort of like wabi sabi.

[Ajmal Aksar @AjmalAksar]

@qiweiiiy @ArtificialAnlys subscriptions change the math a lot. i track my claude code usage and the equivalent api cost for the same work amuses me sometimes. cost per task on a plan is very low, so api cost per task isn't the number that matters for a daily driver

[Ajmal Aksar @AjmalAksar]

@athenaadsops read-only first is the right call. we're building @kAInet_tech as more than prompts. it plans, analyses and executes while you sleep, but you approve every major decision before anything spends. trust comes from the approval step, not from the agent being smart

[Ajmal Aksar @AjmalAksar]

@thenamanraj ides were layers for understanding and navigating code we handwrote, and they earned their place. now that we direct agents instead of writing the code, we becomes managers than programmers, and the ide matters less. so we drift back to the terminal

[Ajmal Aksar @AjmalAksar]

@ClaudeDevs @AgenticAIFdn Will the roadmap include a standard for benchmarking and evaluating skills, so we can tell whether one works the same across agents?

[Ajmal Aksar @AjmalAksar]

@jyriso same path here. long time cursor user, moved to claude and the back and forth dropped a lot. pro's token limits hit fast at first, but what got me to max 20x was running lots of tasks in parallel and experimenting a lot more

[Ajmal Aksar @AjmalAksar]

@svpino with ai, mediocre is the baseline. taste is the nuance you bring. loc jumped, but with clear instructions i've seen codebases get cleaner. people who think in architecture now scale what took big teams. the skill is directing agents well

[Ajmal Aksar @AjmalAksar]

@ArtificialAnlys small models getting this literate is cool and a bit scary. cool because cheap models now get more topics. scary because haiku-priced spam is about to get a lot more readable. curious to see if we can make it write copy with taste

[Ajmal Aksar @AjmalAksar]

@claudeai checks with you before it acts is the right call, it's the same rule we built our ad agent around. just wish it checked with me before going US only

[Ajmal Aksar @AjmalAksar]

@lucas_builds Terminal for sure. Mostly because it runs fine on low-end machines, while desktop apps tend to crash. And with herdr I can run several agents from different tools side by side, which works much better for me.

[Ajmal Aksar @AjmalAksar]

@Al_Grigor The prompt that grows with every fix is what pushed us to evals. We replay real conversations from the platform before each release, and an old error coming back counts as a regression. That made it safe to split one big instruction block into pieces each step pulls as needed.

[Ajmal Aksar @AjmalAksar]

@mrdbourke @ExaAILabs Cost per task is the right unit, since fewer searches per task is where Dynamic Highlights saves. For our agent the other cost is pages that won't load, as retail sites block even clean requests, so we chain about 3 vendors with Exa. Does the index count failed fetches?

[Ajmal Aksar @AjmalAksar]

@mattpocockuk

[Ajmal Aksar @AjmalAksar]

@lostinazeroy yeah, on small fixes the advisor mostly second-guesses things that were fine. next i'm running it on a longer multi-file task, which is where i'd expect it to earn its cost. will share the numbers either way

[Ajmal Aksar @AjmalAksar]

@_ryanskidmore @Cloudflare The part I'll steal is splitting the drop at every release. Only about $0.20 of yours was the cheaper token price, the rest was fewer tokens. Our campaign brief agent is down about 63% per response from caching plus a newer model, and I haven't split it yet. Doing that next.