Mayank Raj (@Mayank9856) header
Mayank Raj (@Mayank9856) profile picture

Staff Engineer @ Stripe | Ex-Salesforce |Sometimes ML, AI, BigData. Big on Serverless. Built https://xitter.gq/t.co/qSkvseB8Yt, building https://xitter.gq/t.co/jX1OfcIYnS in public

Joined December 2014
Photo by Mayank Raj (@Mayank9856)Photo by Mayank Raj (@Mayank9856)Photo by Mayank Raj (@Mayank9856)Photo by Mayank Raj (@Mayank9856)Photo by Mayank Raj (@Mayank9856)Photo by Mayank Raj (@Mayank9856)Photo by Mayank Raj (@Mayank9856)Photo by Mayank Raj (@Mayank9856)Photo by Mayank Raj (@Mayank9856)Photo by Mayank Raj (@Mayank9856)
📢 Announcement 📢 I'll be presenting at the AWS Community India. The session is JAM packed with amazing speakers and I'm excited to share the space (virtually) with none other than @jeffbarr !! More details to follow... #aws #awscloud #Serverless
Speaker Announcement 🚀 Technical Session: Serverless Cron with Second Level Schedule and Conditional flow by Mayank Raj For more details and registration, visit lnkd.in/gDTQgEW #awsugindia #acdindia #aws #awscloud #awsusergroups #communityday #serverless #cronjobs
8
6
31
Insane how good Fable 5.2 is !! Just 17 hours, 47 agents and it was able to rewite Rust in Rust !!!! 17% faster runs, 47% smaller builds
1
50
Most of us have implemented from a static mock and guessed what happens on click Cisco's PubHub team lost flows and animations in the Figma round trip. They refined the HTML instead and got running pages, a persona switcher and a changelog blogs.cisco.com/cisco-devnet…
2
41
1/n A new hire may copy the worst file in the repo, not because they ignored the guidelines, but because it was the closest example. IBM's Palette team saw a similar problem when a teacher model generated training decks.
1
27
2/n In their workflow, the model copied its reference decks more than it followed prompt instructions. Adding prompt rules rarely stopped it from making things up.
1
20
3/n Their fix was to edit the reference deck. In this workflow, changing the example shaped the teacher's output more reliably than adding rules. The example was part of the instruction, whether or not the prompt said so. huggingface.co/blog/ibm-rese…
15
Shopping chatbot is fine until something goes wrong, then you want the attendant Well. Who do you get !? Places where were're judged (comparing features) - prefer AI. When it comes to money - god old human ! baymard.com/research-article…
1
45
Finding the latest value in a long document is one job. Tracking it as it changes is another. In Ai2's StateBench, four frontier models did well at the first. When variables copied each other, the longest sequences dropped them to 13% or less. huggingface.co/blog/allenai/…
26
An empty control plate with the wrong antibiotic is a green test run against the wrong environment. In one BioBasicsBench task, GPT-5.6 Sol trusted the blank plate. GPT-6 Astra read the rim, saw kanamycin instead of ampicillin, and rejected it. openai.com/index/measuring-h…
1
1
93
Normally the CAD file is the original and the part is the copy. Lose the file and it flips. OpenAI + Chip Ganassi Racing are testing the old-school way back: 3D scan -> editable CAD & manufacturing drawings, reviewed by CGR engineers. openai.com/index/acceleratin…
2
125
Every old .bashrc has an alias nobody dares delete ! AGENTS.md is growing its very own set of "NEVER CREATE ANOTHER utils DIRECTORY" Sounds excessive until you've seen utils/, utils2/ and helpers/ Which prompts another line of..
2
2
107
Every non-goal needs a best-before date Say, "No live updates for v1. Ops plans tomorrow's staffing from this, so hourly is enough. Revisit if we assign work mid-shift." The reason says what the data is for The last line says when the no expires
38
All-typed up. Well-documented. Invoice API time's out, returns [ ] "Couldn't fetch the invoices" and "the customer has no invoices" are now the same answer. Some might call a better response..
115
"Works on my machine" has ended more race-condition investigations than any lock. An agent that gets the fix wrong but hands you a loop that fails on demand has done the hard part.
34
1/n Send an exterminator into an old house for one specific mouse, and they'll trap a different one first. AWS Continuum hit 89% on CyberGym-E2E, yet fixed the target bug only 37.8% of the time. In the other 51.2%, it exploited and patched a different C/C++ bug.
1
85
Data pipelines spend millions validating rows, when the quietest corruption comes from a bad join Merge two "John Does" with one's address, the other's loan. Every row is valid. We just gave birth to an overachiever.
46
1/n The average of "fine" and "on fire" is "slightly warm", and no user has ever experienced slightly warm... Say A's p99 is 10ms and B's is 1,000ms. The dashboard shows 505ms. Combined over 10,000 + 100 requests, the real p99 is actually 10ms...
1
52
2/n Weighting by request count still gives 19.8ms. Percentiles throw the distribution away. Merge compatible histograms first, then estimate. prometheus.io/docs/practices…
23