Ramsri Goutham Golla
@ramsri_goutham
Followers
13K
Following
13K
Media
2K
Statuses
8K
Lead Data Scientist at https://t.co/v4OH4htiGp 👨💻 Also a founder bootstrapping 2 AI SaaS Apps to $100k ARR with no employees: https://t.co/fU8yoiYnNE https://t.co/DTyILlia5O
Learn Practical AI for free 👉
Joined December 2017
Take your side projects seriously!. You might as well get featured in the Google I/O keynote on the main stage in California ;). I'm excited to share that the work @ravithejads and I did in building Navarasa, an Indic instruction-finetuned model on top of Google's Gemma catering
76
39
809
Life update - We bought a new home! . After 15 months of planning—from the idea of buying a home, arranging the finances for the down payment, securing a loan, and getting the interiors done to our liking—we are finally here!. After living in multiple rented homes and dealing
337
140
4K
A tutorial from @OpenAI on how to build a question-answering system on your custom website. Preprocess:.1. Website -> text (Scraping).2. Text -> chunks -> embed. Inference:.1. Question -> embed.2. Get topK similar chunks .3. Answer generation using 1 & 2.
26
173
971
The serverless GPU space is heating up with image and text generation startups blowing up 🚀. Here is a map of the startups (mostly unknown) offering serverless GPUs:.🧵.
37
142
913
I will summarize GPT-4 for you!. It basically made @karpathy 's 10-year-old dream come true with the new multimodal (image + text) capabilities. If you give GPT-4 the below picture and ask it to describe it, it can do that accurately while also describing the humor going on
46
100
795
Promptify - A new library for GPT-3 based prompts to get structured outputs (list/dictionary). 1. Coming up with optimized prompts for common tasks like NER, Ques Answering, etc is not trivial. 2. Parsing prompt outputs with regex is cumbersome!. ->
21
99
756
The most practical open-source competitor to @OpenAI 's GPT-3 is Google's Flan-T5. Here are 5 Flan-T5 resources to try out easily, deploy, or fine-tune it!. 🧵.
24
115
735
Most people don't understand @OpenAI GPT-3's tokenization and how expensive/inefficient it is to build a GPT-3 app in non-English!. English: John - 1 token.Telugu: స్తి - 12 tokens 🤯. Byte pair encoding divides a character like స్తి further!. స్తి (12 tokens)= స్త + ి (8+4)
30
66
577
With Descript's announcement of a $50 million raise from OpenAI yesterday and with accurate speech-to-text models like Whisper, the AI audio space is hot right now 🚀. Here is a detailed list of startups, some brand new, in the AI audio/video space:.
28
79
567
I am still in awe of @OpenAI's Whisper!. I am subtitling my courses (highly technical + Indian accent) and in a 20-minute video, I sometimes have < 10 total word edits. Just 6 months back this was a very distant dream! .Now we have whisper released and we built our SaaS
11
44
537
$40k ARR.Solo, bootstrapped, part-time!. In terms of Indian standards, this is like having a well-paid, full-time job that I can't be fired from :)
56
8
543
Want to easily deploy @OpenAI GPT-3 APIs or @LangChainAI apps?. Here are 7 resources to do the same:.🧵.
19
66
500
If you are a full-stack developer there was never a better time to build apps!. Finding ideas:.1. Just got to @promptbase , filter by top-selling prompts (GPT-3, ChatGPT, Stable Diffusion). 2. Buy a few interesting ones. Convert them into a SaaS. Eg 1: .Buy "meeting to minutes"
23
63
465
Grammarly doesn't have an API yet for grammar correction. But we have the newest release Gramformer to offset that by @prithivida !. Github: #NLProc #MachineLearning #ArtificialIntelligence
10
92
454
What's new in the @OpenAI Whisper's world?. Less than 24 hrs ago OpenAI team released the new Whisper large-v2 model. 1. The "large-v2" model is trained for more epochs with regularization and shows improved performance compared to the previous large. 2. Same architecture as V1
10
60
462
I wish OpenAI had a better understanding of what it takes to build a product on top of their API!. 1. I can build on gpt-3.5-turbo-X but it will be upgraded/deprecated every 3 months! So I need to constantly check that all my prompts work as expected every 3 months? A bit of a
53
45
422
Did the impossible!. Personalized guided meditation with AI!. ChatGPT (text) + elevanlabs .io (audio) + dreamstudio .com (images) + supertranslate .ai (subtitles). Asked to generate guided meditation for me who is a bootstrapper building AI products!
23
66
402
Anthropic (Claude and Claude 2) started rolling out access to Chat and API users all over the world!. This means you can start building production apps on top of it!. For context the relative pricing for your understanding:. GPT4-32k context:.60$/million input tokens
15
43
404
🚀 My inspirational journey from a small town to driving a Porsche to work every day at 27 in the Silicon Valley was covered by @thebetterindia Facebook page yesterday. Link:
14
23
373
It feels magical to have created a yearly value of $100k in this world, sitting in Hyderabad, India!. An idea -> Code in a Google Colab Notebook -> Businesses in the real world serving 500+ monthly customers!. No funding! No full-time employees! No ads! .No marketing budget!. It
43
13
376
I am putting together a list of platforms that you can explore if you are looking for economical AI/ML training and inference -. 1. banana .dev - Ship ML to Prod, instantly. Scalable inference hosting for your machine learning models on serverless GPUs. Thread ->.
16
84
356
Watched Jailer movie yesterday!. A housefull show in Telugu at 11 pm on a weekday!. For non-Indians it would be unfathomable that a 72 year old hero can keep the box office of 1.4 Billion people under his hukum (order)!. Anirudh's BGM is top notch and he hits it out of the park!.
6
158
351
In Sept 2022 @OpenAI released Whisper, the world's most accurate speech recognition (ASR) that can transcribe and translate speech audio from 97 languages!. But it had a few shortcomings!. Here are the top 7 community enhancements on top of Whisper that you should know about:.🧵.
18
54
353
Time to give back :). Releasing for FREE on Youtube my 12+ hour course "Practical Introduction to Natural Language Processing". A lot of folks asked about the best resources to learn all the essentials necessary to be ready for the LLM, RAG, and AI Agents era!. This is the only
14
40
352
How do you deploy large ML APIs (eg: text generation) for fast inference with GPUs but without breaking the bank?. Answer: Serverless GPU inference. Here are 5 new age startups solving for serverless GPU inference :. Thread ->.
9
39
314
$90k ARR! . Making money on the internet still feels like a fairy tale! . Bootstrapped, no employees, no one full time and only fractional founders. I close my eyes and just wish that it lasts a lifetime :)
29
6
291
If you have no clue, learn and master.LLM -> RAG -> AI Agents .in that order!.
6
10
290
Want to learn LLM prompt engineering from basic to advanced in a single place?. Check out:
4
44
287
Just hit 2000$ MRR 🚀. Solo, bootstrapped, and a fractional work side project!. 85% profit margin. No affiliate pay. Growth via SEO. No recurring cost of customer acquisition (paid ads etc). "Text in and Text out. Just an OpenAI layer. I can build it in an hour with better UI.
36
8
263
I'm excited to announce that I have joined the Knownwell .com team as a Lead Data Scientist, full-time!. As many of you have been following my journey for years, I will answer all your questions with full transparency below :). But first, what impressed me about the Knownwell
59
7
257
Without a doubt, 2024 has been the best year yet!. Here are my top 5 memorable events:. 1. Got featured in Google I/O 2024 for the work we did on Telugu LLM Labs! . 2. My SaaS apps reached the $100k Annual Recurring Revenue mark. 3. Took a full-time remote job at Knownwell .com
16
4
252
After 4 years of working from home, started working from a coworking space this week!. I am not a productivity junkie but with a kid at home, I forgot what deep work is for the last 4 years. Barely 15 mins passes by before a delivery guy ringing the bell or some other internal
27
5
252
When you are in college the next goal is to get a job. When you are in a job the next goal is to get a top company job. When you are in a top company job the next goal is to get a top-paying overseas job. When you are in a top-paying overseas job the next goal is to make the.
6
32
244
Introducing Telugu LLM Labs!. An independent joint initiative by @ravithejads and me to advance the Telugu NLP space and give a better AI experience for 100 million Telugu speakers in the world!. Link: Goal:.> Contribute to open datasets focusing on
27
22
224
My AI creator journey-. 1. Learnt NLP.2. Freelanced.3. Picked a niche (edtech).4. Released Questgen on Github.5. Questgen into a course (8k$).6. Medium Top Writer in AI.7. Taught NLP cohort course (6k$).8. Learnt SEO and Bubble (no-code builder).9. Built Questgen app (1k $ MRR).
9
18
202
It took me 5 years, but in the end, I learned to:. ✅ Build a startup without capital.✅ Build a startup without co-founders or employees.✅ Market a startup without ads.✅ Earn a living without an employer. And above all -. ✅ I learned to build a startup without burnout and
24
3
213
Tip:.If you are building on OpenAI + Pinecone stack, pick 'dotproduct' as your metric instead of 'cosine' to get minor speed gains while creating the Pinecone index. Since OpenAI Ada embeddings are normalized to 1 anyways, cosine and dotproduct will yield the same results. So
7
25
197
Introducing Questgen- A simple to use NLP library for state-of-the-art Question Generation. Generate MCQs, Boolean (Yes/No), FAQs, paraphrase, and more. Built with #T5 and @huggingface. Check out our submission on @madewithml #NLProc #NLP #edtech .
6
53
187
Officially now I operate three profitable AI SaaS apps Questgen .ai, Supermeme .ai & Supertranslate .ai. GPT-3, Whisper, CLIP & vector search are in production across my apps!. Not bad! I have placed some of the biggest advancements of AI in the hands of general consumers :).
18
16
184
Life is good :). First 10k$ post tax month sitting in India! Grateful for the opportunities!. 60% through full time job and the rest two side projects (Supermeme .ai and Questgen .ai) filling the rest! . Thanks Pat for the mention!.
This side project pays his rent and only requires 1 hour/week to maintain.
13
7
184
🚀 Introducing NavaRasa - An open Indic Gemma 7B & 2B Instructed tuned model with nearly half a million instruction data points across 9 Indic languages. You can try the inference for yourself on a T4 Google Colab Instance: Also, check out community RAG
18
22
185
Trained a LLama 2 Tokenizer in Telugu (Indic Language) to output token count on par or better than English!. So "నా పేరు రామశ్రీ" (My name is Ramsri) comes to just 5 tokens as opposed to GPT-3 that comes to 41 tokens. Let's see if I can train Llama 2 7B with this tokenizer and
13
8
180
I wish there was a real competition to OpenAI’s text generation API!. I pay ~ 800-1000$ monthly across my apps to OpenAI. I have tried all the other players cohere, A121labs etc even for fine tuning! . Can confirm with certainty that there is a lot to catch up on quality!.
25
6
181
Within a month of @OpenAI 's Whisper launch, we had an internal system built with word-level timestamps and speaker identification!. Having previously worked with a startup that does gene sequence alignment in Bioinformatics, I knew we could merge low-quality but
9
21
176
The serverless GPU space is heating up with image and text generation startups blowing up 🚀. Releasing V2 as few companies have shut down (inferrd) and a few renamed (slai .io -> beam .cloud) since V1. Here is the landscape of serverless GPU startups in 2023:.🧵.
5
28
175
⚡ Been wanting to learn Pytorch but the content is scattered everywhere?. Microsoft has a Pytorch course that is divided into 4 neat modules :. ⮞ Intro to Pytorch.⮞ Intro to CV with Pytorch.⮞ Intro to NLP with Pytorch.⮞ Intro to Audio with Pytorch.
4
42
171
🇺🇸 LLM Engineer Interview Prep 2024:. Fine-tune Llama 2 using a new SFT dataset across various domains, such as medical and law. 🇮🇳 LLM Engineer Interview Prep 2024:. 1. Learn to train a tokenizer for Indic languages, incorporating it with the existing tokenizer of the LLM to.
5
10
168
NLP project Idea: Tweet shortener. Problem:.Twitter offers a maximum character limit of 280. Often times it is annoying and needs extra creativity to shorten your tweet if you are copy-pasting a pre-written tweet from your notes or google doc etc. 🧵. #NLProc #MachineLearning.
12
28
165
Pathway to making first passive income $ as a data scientist -. 1) Colab Code -> API (using Serverless).2) Sell API -> @Rapid_API .3) Landing page for your niche APIs.4a) Build a product without code @bubble or .4b) With code and host on @vercel and @supabase for free.5) SEO.
5
19
161
Here is how we bootstrapped 3 AI startups with positive unit economics -. 1. Development - Google Colab.2. Inference - serverless GPU providers.3. AI Backend logic - AWS Lambdas.3. Semantic Search - Free to start vector DBs (eg: pinecone .io).4. Deployment - Vercel + Supabase.
2
26
160
With accurate speech-to-text models like @OpenAI's Whisper and single-shot translation to English, AI-automated dubbing is poised to become mainstream. Here are 12 companies leading the AI automated dubbing space ->.🧵.
8
18
150
After almost two years of SaaS journey, finally, I am at a comfortable place to not worry about next month's finances as my apps hit $6k MRR combined!. Otherwise, it was a constant jugglery of making courses, doing part-time consulting, and funding my SaaS projects with time and.
23
10
144
Introducing MAGIC AKA "text to GIF" at Supermeme .ai 🚀. Enter any text like a tweet, blog post intro, or your chat reply to a friend, and get relevant GIFs along with funny overlayed text on it! . Just signup at Supermeme .ai and try it out today!. Wouldn't be possible without
8
24
141
Want to build full-stack apps purely in Python without knowing any front-end technologies?. Here are 4 startups leading the way:.🧵.
11
31
137
🚀 Introducing Telugu Gemma 7B Lora Finetuned Model - Perhaps the best open LLM model yet for Telugu text generation out there!. What's a better way to witness the results firsthand than experimenting with it yourself in Google Colab (on a standard T4 instance)?. Inference code
13
22
133
Learning full-stack development from this YT channel!.Just six videos and each one is a masterpiece!. Videos on building Spotify, Airbnb, Twitter, Netflix, etc on a modern tech stack (NextJS 13) with the backend!
4
13
130
Want to build full-stack apps purely in Python without knowing any front-end technologies?. Here are 3 startups leading the way -.
10
26
127
AI, ML, DL, Data Science, Computer Vision, and NLP explained for beginners. My simple to follow guide - .
1
27
122
This month I published 3 unique AI projects with easy Colab Notebooks. Few shot text generation with T5. Semantic search with OpenAI CLIP. Generate distractors for MCQ. #NLProc #MachineLearning #DataScience
3
35
122
The tech stack that we used to bootstrap 3 AI startups with positive unit economics from the beginning - . 1. Development - Google Colab & Internal Streamlit tools with included hosting from @DatabuttonHQ . 2. Inference - serverless GPU providers with pay-per-use model eg:.
4
24
122
Took break from programming and AI updates and celebrated our kid’s 2nd birthday over the weekend 😀. Car entry with smoke 🎉
8
5
123
Here is how we profitably bootstrapped 3 AI SaaS apps to 2750$+ MRR -. 1. Development - Google Colab.2. Inference - Serverless GPU providers.3. AI Backend logic - AWS Lambdas.3. Semantic Search - Free to start vector DBs (eg: pinecone .io).4. Deployment - Vercel + Supabase.
3
6
119
How on planet earth can a word like స్త్రీ (woman in Telugu) come upto 18 tokens in @OpenAI 's GPT-3 whereas a woman in English is 1 token?. This means it is 18x expensive as well as 18x slower because GPT-3 needs to generate 18 tokens for one word in Telugu here!. Deep dive 🧵
13
15
121
My portfolio of 3 AI SaaS apps reached 4500$ MRR :). Text to Meme - Supermeme .ai.Text to Quiz - Questgen .ai.Video to subtitles - Supertranslate .ai.
10
5
113
GPT-4 Turbo with vision is generally available!. I built an image to quiz creator where one can upload any snapshot from a textbook or newspaper and create an instant quiz to practice :). It is a premium feature live at: Supports quiz creation in various
9
6
112
Supermeme just reached 500k signups!. $0 marketing spend!. Fully bootstrapped, fractional work, and a team that has never met each other in real life!. Bloody sweet 🍒 :)
14
1
110
Questgen reached 1000$ MRR 🚀.Solo and bootstrapped!. I am not gonna lie! . It has been a tough battle against the family's constant push to take on a full-time job and also handling the stress of raising a toddler!. But here I am! 🔥
22
4
107
In the pursuit of avoiding depreciating assets, we forget that life itself is the ultimate depreciating asset!
5
4
107
How they announce their new LLM models: . Mistral AI Vs Rest of them 😅
3
7
102
Try out live, the demo for state-of-the-art MCQ (Multiple Choice Questions) generator from any content built using T5 transformer, @huggingface, and Sense2vec. #NLProc #nlp #t5 #huggingface #edtech.
4
34
105
My wife once showed a celebrity Instagram photo with this view!. Gifted it to her today :). Perhaps the best view from a room in this world! . @khyberhimalayan
11
0
103
PPP (Purchase Power Parity) for LLMs, please :). Can't be paying 36x the price to express 'complex' in another language!
13
10
101
Took the famous "success kid" meme and generated variations with ControlNet and Stable diffusion! . Changed race and gender when compared to the original meme!. Might add it to our app @supermeme_ai in the future :)
12
4
98
After 4 years since last digital purchase, upgraded to MacBook Pro M2 and iPhone 14 Pro Max ! . Still holding the title of poor Indie hacker among relatives :)
11
1
99
This year's proudest achievement!. Bootstrapped 2 AI SaaS to 14000$ + MRR while doing AI consulting, selling courses, and raising a 2-year-old!. And the third SaaS supertranslate .ai just began its MRR counter :)
7
2
98
The pgvector support in @supabase is magic 🚀. You can build lightweight similarity search applications just on your DB and without any vector databases!.
4
10
94
There hasn't been a better time to build full-stack apps!. 1. Nocode frontend + nocode backend : @bubble .2. Nocode frontend + code backend : @weweb_io + @supabase .3. Pure python frontend + backend: @pynecone_io or @anvil_works .4. Pure Python Internal app: @streamlit.
6
22
94
My journey with AI MicroSaaS apps-. Questgen .ai (15k users) - Text.Supermeme .ai (100k users)- Images .Supertranslate .ai (1k+ in Beta launch) - Audio & Video. I have now officially built production AI SaaS from idea to execution on all modalities (text, image, audio & vid)😀.
14
0
94
In Sept 2022 OpenAI released Whisper, the world's most accurate speech recognition (ASR) that can transcribe and translate speech audio from 97 languages!. But it had a few shortcomings like a lack of word-level timestamps and speaker identification!. Here
4
14
89
From prompt engineering -> OpenAI's Whisper -> Serverless GPUs -> Python full-stack apps -> AI applications, I discussed quite a few topics this month. My top 6 threads this month (Jan 2023)🧵.
1
11
89
The fall of BYJUs is not just some causal VC funded startup collapse!. It is a restoration of faith in humanity and karma!. What BYJUs have committed is equivalent to modern day crimes against humanity robbing the poor of their livelihood in the hope of a better education to
7
12
87
Two months back Supermeme .ai was at $3k MRR and we paid a $1k Google Translate bill and $800 OpenAI bill. Got some mocking remarks on our margins for being an Ai startup!. Today we are near $5k MRR, switched to Microsoft Translate (2x price reduction), and split traffic between.
4
3
85
Officially now I operate three $ making AI SaaS apps Questgen .ai, Supermeme .ai & Supertranslate .ai. CLIP, GPT-3, Whisper & semantic search are in production across my apps!. Not bad! I have placed some of the biggest advancements in AI in the hands of general consumers :).
6
7
86
If you are building in AI, serverless GPUs are your best companion!. Here is my article listing the serverless GPU providers in 2023!.
3
16
87
Pro Tip:. Are you using Pytorch Lightning in your project but setting your seed everywhere in NumPy, random, and torch?. Use this instead ->.import pytorch_lightning as pl.pl.seed_everything(42). #MachineLearning
6
11
85
Tutorial coming tomorrow!. Create descript like audiograms for free using Python and OpenAI's Whisper from any video/audio!
7
8
83
With all due respect to the open-source contributions, the inference API of @huggingface is robbery in broad daylight. Here is my story using it and paying 750$ for CPU inference which could cost me 300$ for self-hosted GPU inference. CC: @julien_c and @ClementDelangue
11
9
84
Just for the laughs :) . Summarizing @OpenAI 's plugin announcement in two memes! Brilliant move!. Love them both OpenAI and @LangChainAI !
6
4
87
To make $100k from an employer, you need to master a skill. To make $100k without an employer, you need to master the world!.- Nostradamus (1503-1566). Glad to be back at the $100K mark after churn and recovery over the last 2 months :)
7
1
85
4k MRR - Solo, bootstrapped, and part-time! 😃. I ain't a rapper, just a GPT wrapper!.Ain't got a moat, but might soon cop a boat!.Just sailed past four grand MRR!.SaaS game's strong, en route to the success of RRR!
9
1
84
If you missed it earlier, we released an end-to-end tutorial to programmatically generate Youtube shorts in 1-click with cutting-edge generative AI tech!. Full code and a ready-to-use Google Colab notebook are also included. Do check it out:
5
12
82
It is incredible that @_skris built an intuitive editor for our app based on OpenAI's whisper, casually over a few weekends. Now you can subtitle any video to English, split, merge & adjust timecodes and walk away with a final subtitle file. Try for free:
5
15
84
Someone graduating from college and interested in entrepreneurship asked for a one-line piece of advice!. I told him to "Live physically in India, digitally in the USA, imaginatively in 2035!". 1. Living physically in India. India is still cheap! It accommodates a $100 per month.
5
7
83
New video dropping tomorrow in the Practical LLMs tutorial series! . We will adapt the Gemma 7B Modal (without fine-tuning) for classification by modifying the last linear layer to map to only a subset of 3 tokens (positive, negative, neutral) from 256k tokenizer so that we do
4
2
84