AMA 🎙️: Hi, I am diffusedbandit, Senior ML Engineer at an Indian Unicorn
Hey everyone, happy to answer any questions around Data Science/ML/AI and making a career within it A brief about me, I work currently as a Sr. ML Engineer at an Indian unicorn startup I graduated 5+ years ago and have since worked as a Data Scientist / ML Engineer at an ecommerce company and social media company. My designations have changed based on the industry requirements and work too changed based on that. Would love to take questions on what it means to be an ML Engineer, my early learnings from a career in data science, and thoughts on AI. Look forward to all questions :)
1. How much is your pay currently?
2. Did your ESOPs amount to anything?
3. Your highest degree? Thoughts on Bachelors/Masters/PhD. and their impact in this field?
4. Thoughts on Career Progression within the ML/AI space.
5. Ever wondered to make lateral moves? ➡️ Quant/Analyst/Consulting etc.?
1. 60 LPA
2. What companies in india have IPOd? I have ESOPs but I value them as 0. Invest in parag parekh flexi cap folks for wealth creation!
3& 4. No lateral moves, I see myself as a product owner (owning not just ml) in the future. I absolutely love the kind of shit companies around me do.
BengaluruLandLord
Stealth
a year ago
A simple exploration on growws mutual fund comparison reveals quant active and even quant tax plan have consistently outperformed parag Parikh by a long margin
See more comments
thanks for doing AMA,
1. What does a typical day look like?
2. What are the most exciting recent developments or applications of AI that you find particularly interesting or promising?
With seniority, part of my day goes into meetings - discussions around what is it that we are building, progress checks, brainstorming around what to build next, PRD etc. I also have the added on responsibility of a team lead, where I have to delegate work and ensure timelines are met - in other words - do standups and unblock people. I try to dedicate the rest of the time for coding purposes. Some days hacking things together in a jupyter notebook, other times deploying some change, some days integrating the ML stuff with our backends.
Again, not many people have the role I do, where I have to and can make contributions to model building, production deployment and planning. So the above stuff varies for others around me where some do more hardcore model building & analysis
On question 2,
Gen AI is truly a game changer, ChatGPT, Copilot & the likes enable everyone to ship faster. Designers around me use the Adobe beta image gen functions for their work. I have seen some backend folks generate code from tools like locify from a figma mockup.
Rewording one of the popular sentiments around the recent developments - Gen AI enables the experts to be more efficient.
See more comments
CuriousCroissant
Student
a year ago
As a fresher graduating from college, preparing for ML roles, what specific areas should one focus on especially to crack an interview?
And how does does DSA play a role in interview?
Strangely DSA is something that is never ever useful but rated super highly in your interviews. I would recommend for freshers to continue the leetcode bullshit, and build side projects around to showcase your ML capabilities. Get on twitter and start getting inspired by indie developers hacking things together.
Any company looking at your online work would love to have you, as this separates you from the 1000s of others who also "want" to do AI and fancy shit
CuriousCroissant
Student
a year ago
Ah thats a bit sad...This screening is very worthless I feel and considering the ongoing internships in hand(ML intern @Siemens, AI Product intern at a Startup) it becomes difficult to practice DSA(and frankly I don't like it I'd rather want to experiment with a new model). I've even won Smart India Hackathon'22 and also participated in many other hackathons have other hands on experiences too but just the notion that you've to go through that DSA round despite dev skills crushes me. Can't complain much , have to learn it the hard way :(
Still hopeful and working hard 💪
Thankyou.
Azreal
Stealth
a year ago
How do you manage to stay upto date on the new emerging technologies and updates in the filed ?
Asking as I'm currently facing that issue as a fresher(1yoe))
Twitter!! I can share later the accounts you have to follow to be uptodate. But that's the only platform you have to be on to be aware of what is happening.
See more comments
What do you make of the current generative AI wave?
Is it a bubble or is it the future?
Definetly a wave, in the context that many companies will come and get popular and die right away.
But the difference this time is, Gen AI is backed into everything now, because it is that good & useful.
Gen AI for images in Adobe, the features around text generation in grammerly, docs, notion (everywhere).
I do not believe in no code BS tools, and I do not beleive anyone saying write a text and get a website out, this is not how things work at any minimal scale, but a developer can now do work that a team of developers could do a year ago.
Same goes for every other profession that has these advances.
Thanks for that reply @diffusedbandit; glad that your views echo
Curious: So far in your role have there been any changes? Or is it mainly other functions that are getting better because of easier text/image/video generation
See more comments
How much value do you think a masters/PhD would hold in the space. I've been working in the space for a couple of years and am a little confused on how to proceed
I personally would do what Thiel does as soon as I get that rich. Which is to sponsor people to drop out from the colleges/universities. Because they are truly waste of your time if your ambition or goal is something else / bigger.
By goal - if your goal is to do heavy research then yes, phd & masters is the way, you can get yourself into google and the likes and get something massive out. But if your goal is to be the data scientist at udaan, I don't think you need masters or phd for that
Having said that I can see why an employer would prefer a masters & phd candidate over a undergrad with some experience.
Azreal
Stealth
a year ago
What if the person wants to get the international exposure but can't due to not being from tier 1&2 colleges
See more comments
Noctus
Stealth
a year ago
what tech stack are you currently using for ai developement
Everything. My role has me prototype faster. So huggingface abstractions wherever available, tensorflow serving / torch serve (yes that too) for things like encoder model deployments. For usecases with user data (read big data), PySpark.
I have not seen diversity around what thing people use around me. At the end of the day it is still python and some python packages. And then you go to prod, where things change
For what's it worth, does MS in Data Science or applied ML hold weightage in today's market? Quick background - fresher in industry, preparing for Masters this fall - conflicted between distributed systems and data science roles.
different companies, different hiring managers, different philosophies and expectations from the perfect candidate.
I would always let my work get me my next job than my degree, because of my biases.
See more comments
Bot_God
Stealth
a year ago
Are mostly complex models use or is it more depending on task.
How much learning -> Execution takes place
say a user modelling problem, like recommending posts to a user on grapevine - one could make a case that simple recency + affinity based models could work. you could take up LR next to solve it and so on.
then say a text classification task, moderating say grapevine posts, hypothetically, I could see using an LLM based prompt engineering based classification model.
Companies in bangalore I do not see anyone sponsoring your pay to make a foundation model, they need you to execute faster, that usually implies that you won't learn as much but rather apply more and more
Bot_God
Stealth
a year ago
Nice answer. I think mostly execution is needed and accuracy up to +30% Works too, as compared to linear models.
Have you never thought of going to Algorithmic Trading side ? Higher pay is there
See more comments
I have not met many 10 year DS folks. But this question is quite difficult to answer.
Companies like sharechat or walmart have different bands, companies like Infosys have different bands. You might work on the same tech but it really depends where you work.
For a product based company , what one can expect ,? A range could help
Do you guys build any fundamental models/ something that is truly unique? Or the mandate is to build wrappers around stuff that already exists?
HELL NO. Nobody in bangalore apart from MSFT likes is going to sponsor foundation model building.
Do you think Flipkart would let me build an LLM or let me work on a model that recommends the products the user is likely to buy next? Its all Applied AI, wrappers around things or classical predictive models
What is your view about those who do frontend or those who develop meaningful interfaces post ur data crunching ?
very crucial, I believe that the DS/ AI / ML folks get the most glory but the people behind the scenes - platform, production, backend, frontend, without them you are just a python code in a notebook
CuriousCroissant
Student
a year ago
The models which are build from scratch are they built on c++ or python? Coz some from big companies use cpp as its faster
Model building usually happens in the commonly available and mostly used frameworks. Model inference happens on frameworks written in C++ and optimized for that. Just Python can't do cost optimizations, super low latency predicitons alone
CuriousCroissant
Student
a year ago
Ohhhh..nice!
everything
Stealth
a year ago
How can someone with 1 yr exp as sde can transition to a ML Engineer?
Check if you can transition into this role at your current org? That is usually the best option. Also learn up the ML engg skills and showcase your ability that you can do those things via github etc.
transition from SDE to ML Engg role is likely to be easier if you switch to a early stage startup, in my experience, there, people are more comfortable with competence than labels, and the roles are never hardcoded / bucketed
What would interview prep look like for 5 YOE mle?
I assume it would be stats, ML theory + deep learning based questions based on past projects+ ML system design and some Medium DSA on some rounds?
Different companies require different preparation, best to ask straight up in your interview rounds (before I mean) what they expect.
Just a hack btw - senior you become, try doing people management, it would really help you not to be the most perfect in DSA, System Designs, ML, Stats and 1000 other things everyone wants you to be perfect in. Hack around.
Yeah actually, i meant in context of Indiam unicorns which are in e-commerce space for example. Like your team.
I get your point about people management and have been doing it now for 3 people.
1) have you ever thought about or worked on ways to mitigate CUDA supremacy? like experimenting with Habana, AMD hpu.
2) I have seen tech people in my org (with DS background) talks about every buzz word out there..be in LLaMa, Falcon ..Langchain, babyAGI ...yet they can't improve a basic classification accuracy more than 1 to 3% f1 ... do you have such people in your org? what are your thoughts on that ?
3) Being in startup do you face any resource shortage when it comes to training infra and possibly inference infra ?if yes how do you mitigate that and is that encouraged and noticed by higher leadership?
4) where would you prefer to work .. under a well versed technically inclined management with above avg industry pay ....or under absolutely non technical management under FAANG level pay?
1. as a side project that would be absolutely fab to take up. I mean challenging the market leader, the 80%+ share of nvidia? why not. But there's a reason why its not done, which is because its difficult and requires core competency. And lot of love for low level tech which I do not have.
2. All kinds of people exist everywhere. I guess, its really upto whoever wanting that extra % gain for them to hold the folks responsible. DS is hard. People either expect that shit works 100% of the time, and are not happy when it works 98% of the time. Folks either have a full automation approach or no automation approach. This is changing though.
3. Companies I worked with have not wanted to work on training something novel or foundational in nature. So resources have never been that much of a question. Although when big data is concerned, yes, it becomes painful when I can't work on some dataset because the data set is not prepared or takes too long.
4. subjective answer - I would want to build solutions for the business problem I am working on. I will also always work when the tradeoff of pay vs work vs people I work with makes sense. You do you basically.
hey thanks for the reply.
help me understand the 3rd point in detail,
I agree on not making foundation models but still you might have custom data on which you would need to atleast fine-tune. given that you use any HF 7B-13B foundation model to further fine-tune. that it self needs some training infra. 24Gx4 gpu at minimum for decent batch size.
have you face such inhouse fine-tuning requirements? or you prefer to rely on API based solutions (for ex. openai provides API to finetune their davinchi on to your small dataset)
Noctus
Stealth
a year ago
Could you kindly share the important skills you have acquired to become a Sr. machine learning engineer?
Ownership is what separates me from others, also the stupid appetite for picking anything that is non ML just to see my ML work getting used. But its more how you are able to understand business requirements and translate your ML work into impact is what people would value you as a senior
Do you guys rewrite the models/mathematical algos in scala/golang for production?(from python to scala/golang) or make the python models work at production levels
ok here's something that knadh & folks once wrote in their zerodha tech blogs.
rewording, obviously, when your db sucks and can't do more than 1000 QPS, do you really need to have a backend that does 100k QPS?
in other words, we use golang / rust for high qps projects, and python (even the uggliest looking code) for anything that is low QPS
BengaluruLandLord
Stealth
a year ago
1. Do you feel Data Science team is over staffed at your company?
2. Is there enough influx of work now and do you see them coming in future too?
3. What about your seniors(principal DS, or director of DS and such). How outdated are they?
4. How would rate your expertise in building ML software (that would work in real time) vs one analysis and insightful work that would aid business decisions.?
Are you good with pure analytics or just a model development and push to production and maybe monitoring kind of guy?
5. Do you think Data Science has good political power in an organisation? Am talking about how easy/ difficult it is to have your DS folks get budget/promotions/moving up the ladder etc... How much political power does DS as an org command during negotiations? Is it significant or more like waiting for alms from management?
1. Its adequate. Rather understaffed.
2. Work is always going to be there. There are always areas you can improve upon, internal processes you can automate and so on
3. Not outdated. Although repeatedly we have seniors who are out of place, are not able to perform, and they either go or are layed off. Not at super senior positions.
4. I do full stack ml. Circumstances had me learn production grade inference, react frontends too.
5. Very less, sad. But how much political power does tech have? More or less the same power as DS. It depends on the leaders, how much they value data and data driven decisions
BengaluruLandLord
Stealth
a year ago
Really? @diffusedbandit I feel engineering is way too powerful politically. That doesn't exist for DS somehow. We are big too. Think unicorn atleast
See more comments
steppenwolf
Stealth
a year ago
Hi, I am a B.Tech graduate from Tier 3 College. I have 3 years of Non tech experience and currently i am working in a small company that sells data. My Current work consists of Data cleaning, Pre processing and prediction.
I want to move to a Data scientist role where I can work on many complex tasks or build systems. I have learned Classical ML, Neural networks, currently working on NLP and deep learning required for NLP. Can you suggest some tech stack or skills i need to learn that will increase my chances of getting hired, or something thar is chrrently Hot in the industry.
I have good 4 End to End Projects in my resume in modular coding.
Thanks.
Th
changing domains is tricky, requires a bit of patience. Check if you can transition to a DS role within your current org?
If not, build something DS / ML / AI on the side and get people to use it online. If any person hiring looks at that level of competence you will have good chances for that leap of faith required to transition.
Discover More
Curated from across
