enterprise ai and first principles thinking go hand in hand
bengaluru
- Tweets 2,223
- Following 680
- Followers 312
- Likes 8,285
with ai speeding up almost everything, right from poc to build till deal closure, your hiring philosophy should revolve around your enterprise sales cycle, and you should optimise for how can i get in high-agency people at each stages of the sales cycle. 1/3
1
45
pocs are becoming full-fledged apps in matter of few days, i am closing pocs in 3 days, where i act as a solo product builder, taking care of everything end to end, from engagement, build, narrative, and handoff to sales. 2/3
1
36
Atharva Ajit Patkhedkar retweeted
What used to take years to ship has been happening in weeks with AI.
Skepticism from a lot of corridors is echoing that building in consumer-tech is thus futile.
Me and @harshitshukla70 strongly feel the opposite.
20
33
66
1,570
had a lot of fun building this, amazing work with failproof by @niveditjain, @Niki_agi and the team!
artificial (@atharva_ajit)
Evidence Gate for pharma GMP deviation investigations: an "operator error" root cause gets flagged by jev when records show an overdue calibration, and passes when the evidence backs it up. A brand-new domain.
github.com/atharvap1209/jev-…
4
139
we have the model transcribe each page into html rather than json. handwriting comes back as a coloured span, struck and overwritten text as inline annotations, tables as actual tables. so every check after that is reading a document instead of looking at an image. the expensive pass happens once.
63
we were making the agent populate our data model field by field, through tool calls. it felt rigorous. it was also where most of the cost was.
now one vision pass transcribes the page into html and everything downstream reads that. roughly 4x cheaper, and the fidelity held.
2
68
broke down what a single run of our pipeline actually cost. the model thinking wasn't the expensive part.
it was cache writes. re-sending the same context into every subagent, then appending every tool call back into it, over and over.
51
first month we tracked precision and recall per page. the numbers moved and the product didn't feel any better. then we started sorting errors into types instead of counting them. a handful of failure modes accounted for most of it, each one needed a completely different fix.
2
78
i spent days reading traces to find out that the most common failure wasn't a wrong answer. the model kept quietly correcting typos in the source document.
that's a feature in most products, but in document review the typo itself is a finding.
52
in a multi agent system the thing that breaks is almost always the seam between two agents.
four questions per subagent:
- what part of the parent's context crosses over
- what shape comes back
- what counts as done
- how the parent checks done before it accepts
fourth is finessing.
66
the prompt is everything right up until there's a second agent in the system. then you can rewrite it all day and nothing moves. it took me weeks to stop trying to keep editing the prompt.
1
55
i have a front seat to enterprise AI adoption and i love it. will be posting my learnings from last 3 months super soon
2
69