Prakhar Upadhyay

Senior AI/ML Engineer · GenAI builder | Learning inference engineering (vLLM, PagedAttention, speculative decoding) | Open to roles in inference engineering 🚀

[Prakhar Upadhyay @404notacoder]

@ahab_developer Tell me something i don't know bro 😛🤣

[Prakhar Upadhyay @404notacoder]

@MaxToks001 Yes 🙌

[Prakhar Upadhyay @404notacoder]

@eliana_jordan Yes 🙌

[Prakhar Upadhyay @404notacoder]

@RakhechaShubham Yes bro 🙌

[Prakhar Upadhyay @404notacoder]

@ralucaroszkos But it looks cute 😛🙌

[Prakhar Upadhyay @404notacoder]

@RadadiyaSunny U look sexy bro 💪🙌

[Prakhar Upadhyay @404notacoder]

@himanigupta000 Yes becoz it's important i guess😛

[Prakhar Upadhyay @404notacoder]

@darihodl Fuck Sky ...U can go even beyond that 🙌

[Prakhar Upadhyay @404notacoder]

@bowora_com Sonnet 4.6/sonnet 5 is good in json output data structuring.

[Prakhar Upadhyay @404notacoder]

The token arbitrage reality in 2026:
| Model | Input / 1M | Output / 1M |
|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 |
| Gemini 3.8 Flash | $0.75 | $3.75 |
That’s a ~13.3x cost spread.
If your multi-agent architecture is routing basic webhook triage and deterministic JSON parsing to a frontier flagship, you aren't doing AI engineering—you're just subsidizing datacenter cooling bills.
The runtime moat isn't the model. It's intelligent routing. Fix your dispatch layer.
#LLMOps #AIEngineering #MachineLearning #buildinpublic

[Prakhar Upadhyay @404notacoder]

@achyuthangr Yes sir,currently building this https://t.co/naMelgAvlo

[Prakhar Upadhyay @404notacoder]

@shahamalig Sexy tea sir...just like un😛

[Prakhar Upadhyay @404notacoder]

@darihodl It's ok don't cry...I can't see you crying 😅...what are you building btw ..🙌

[Prakhar Upadhyay @404notacoder]

@shahamalig sexy tea like u sir 😅

[Prakhar Upadhyay @404notacoder]

Update from the engine room: Replicate support officially confirmed the platform issue causing predictions to hang indefinitely in the "Starting" state is resolved.

Our inference pipeline for https://t.co/Oag9gfl1Pu is unblocked and back online.

Wrestling with upstream GPU hiccups taught us a crucial lesson: you can't control third-party infrastructure, but you can build a resilient system and an effortless user experience.

Here is what we’re shipping next before our public launch:

1-Click Inference Re-render: If an upstream GPU worker stalls after LoRA training finishes, the customer never has to re-upload photos or wait 20 minutes to retrain. Their AI model weights are preserved, allowing a 1-click instant re-dispatch of all 50 portraits.
Calm, Async-First UX: No raw technical errors, confusing logs, or anxiety-inducing progress bars. Once photos are uploaded, the UI gives one clear message: "Your headshots are being generated. We’ll notify you by email as soon as they're ready to download."
Visual Upload Guidelines: Adding side-by-side visual examples (lighting, angles, expressions) directly into the upload flow to help users submit the best possible training set.
Pre-Flight Face & Blur Detection: Implementing automated pre-screening (face detection, blur analysis, and landmark verification) before images ever hit a GPU. If a photo is blurry, filtered, or has multiple faces, it gets flagged in the browser before compute spend begins.

The core pipeline works, the resilience layer is locked in, and we are dialing in the final UX polish.

We are launching very soon. Let’s fucking go! 🚀⚡

#BuildInPublic #IndieHacker #AI #MachineLearning #GenerativeAI #ReplicateIssueFixed

[Prakhar Upadhyay @404notacoder]

@kasikcipsi Ok let's get married then 😛😅

[Prakhar Upadhyay @404notacoder]

@victoriapota