viig99

Switch workflow orchestration to temporal!

You are defining an exceptional Product Manager not the norm.

RPC generally means server side calls, probably this https://twitter.github.io/finagle/, and XHR is not RPC.

Said engineer, said 1000 RPC calls are not the issue, and when questioned further on proving a tracepath / perf result casually exited the conversation, to me it looks like they are not ready to do better or listen to valid criticism, leading conversations on semantics is just creating noise instead of acting on signals.

Also 1000 RPC's are a very valid performance bottleneck, 1000 * latency caused by serde+ network can compounds easily based on data locality + container network orchestration, no data was provided otherwise even when asked for.

Clearly these engineers have never done customer feedback for developing nations, twitter interactions clearly provide evidence for what musk said, but some engineers seem to be in denial, imho the firing is cause worthy.

Location: Toronto, Canada

Remote: Yes, (Open to Hybrid and in-office)

Willing to relocate: Yes

Technologies: Machine Learning, Pytorch, C++

Résumé/CV: https://viig99.github.io/

Email: accio.arjun@gmail.com

GitHub: https://github.com/viig99

--------------------------------- Applied Machine Learning Engineer with 11+ years experience in backend engineering and machine learning.

I have experience in solving for end to end machine learning problems, robust data collection & cleaning, distributed training on GPU clusters, building custom model & training recipes on domain specific problems, model calibration, scalable and low latency inference pipelines in C++ / python (GRPC, k8), monitoring and hard negative mining, Model specific inference optimizations (Quantization, Pruning, Torch Dynamo etc) and have worked on various domain problems like: Search & Ranking, Cold start recommendations, Constraint based optimization, Speech (Speech to text, text to speech), Vision (Segmentation, Classification, OCR), NLP (Doc QA, Classification, Entity Recognition) etc

https://twittoons.com/26, seems true

TBH high levels skills on C++-20 is not what these companies are looking for, instead on C++-14 or so, meanwhile anyone learning C++ in 2022 would rather want to pickup C++-20 ++ with its niceties, old C++ is too tiresome without extra rewards.

Which is good, he is invested in making it work, much better for the customer compared to whatever the product currently is.

Same, "We knew that C++ was harder to scale and maintain high quality as you build a dev team" this just sounds arbitary and a weak excuse to use rust, C++-20 is as scalable as rust with a very rich ecosystem.

August 15th 1971, President Richard Nixon announces that the United States will no longer convert dollars to gold at a fixed value, effectively ending the Bretton Woods system. He also imposes a 90-day freeze on wages, prices and rents.

Meh, trump's gonna whine democrats stole the election, media is constantly talk about him for next 4 years, he will win again in 2024 and nothing is going to change. Meanwhile climate change, economy and social media nonsensence will keep getting worse, why ? because nobody wants to really fight for the little guy.

Exactly, dont think its that complicated.

Looks pretty neat

Op says he has thousands of customers without any proof whatsoever, how does this work in rains ? power outages, what about customer suport, seems like an half baked pitch.

SymSpell port for python using c++-pybind11, is 3x-40x faster than symspellpy the current best symspell port in python and maintains same api structure.

This is super cool!

pytorch is pybind11 + c++

I personally find C++ + pybind11 vastly easier to work with, also transitioning completely to c++ from there was a pretty small leap.

Yes accuracy, latency & throughput are the 3 poles we try to achieve, c++ helps with latency & throughput and helps keep the cost low.

ML engineer here, team started as a research team, now that we have things in production and have a lot of devops, engineering work, we bifurcated into pods and work on specific bits and pieces, lot of constant fire-fighting though. Re-wrote entire stack from python to C++ threadpool async grpc (is thrift the only good threadpool server implementation available ?), deployed on openshift, used vector + influx + grafana for dashboards / internal model monitors, elastic search for loggings, lot of other tools for validation, filtering for potential training candidates etc. Right now working CI/CD for ml, during training if model finds a better model based on different validation sets, have one click deployment ready for approval etc

Winter is here, ah damn.

This + remote work is a boon from gods, i will happily go enjoy life in the wilds.

Staggering over-generalization

Fast C++ implementation of ESOLA using KFRLib, can be used for online time-stretch augmentation during SpeechToText training.

Modern c++ is surprisingly easy to pickup especially for developers coming from other languages, very easy to focus on the good parts and totally disregard the messy bits, working with older libraries is still a pain though.

https://github.com/daptin/daptin, DIY self hosted firebase and more!

Lee sedol is a pretty cool guy and a normal human being, albeit the crazy ability to remember games likes maps, do watch the latest episode of 'master in the house' a korean variety show which featured lee sedol as the guest.

Scheduling is a knapsack problem.

Bert has an 15% masking rate, seems co-related, also 90% is what works well when you are trying to do label smoothing using entropy minimisation, what's going on!