https://arxiv.org/pdf/2402.17152 we have new kind of Recommendation engines thanks to transformers and next token prediction, we can predict the next token and that token don’t need to be just text or character it can be anything!
llm based recommendation engine with traceloop/telmetry kind of dx support, and posthog kind of ux would be a really killer feature
clevertap is great example when it comes to product market fit and what peak great example looks like
need this level of dx: https://www.tinybird.co/
tiktok recommendation algo breakdown: https://www.youtube.com/watch?v=8WVthsOAJgM https://arxiv.org/pdf/2209.07663
nvidia doing related stuff: https://github.com/NVIDIA-Merlin/Transformers4Rec https://github.com/NVIDIA-Merlin/Merlin https://nvidia-merlin.github.io/Merlin/stable/index.html
https://research.unboxai.com/foundation-model-for-consumption-transactions-and-actions
companies which are using it in the production: stripe: https://x.com/thegautam/status/1920198569308664169 pinterest: https://www.youtube.com/watch?v=XdAWgO11zuk linkedin: https://www.youtube.com/watch?v=U0S6CfzAY5c&list=TLPQMTcwNzIwMjW6t1J2h4S2Yw&index=2
netflix: https://www.youtube.com/watch?v=AbZ4IYGbfpQ&list=TLPQMTcwNzIwMjW6t1J2h4S2Yw&index=3&pp=gAQBiAQB instacart: https://www.youtube.com/watch?v=PjaVHm_3Ljg&list=TLPQMTcwNzIwMjW6t1J2h4S2Yw&index=4
gemini: https://www.youtube.com/watch?v=LxQsQ3vZDqo
BETTER GENERALIZATION WITH SEMANTIC IDS: A CASE STUDY IN RANKING FOR RECOMMENDATIONS: https://arxiv.org/pdf/2306.08121
lora adapter kind of mechanism is kind of cool, similar something concept can be used here also, where there is one big modal capable of doing everything in generalised way, but if you want to extend it or make it some specific you can do t very easily!
caption generator: https://huggingface.co/Salesforce/blip-image-captioning-large
fast vector generator: https://github.com/tensorchord/VectorChord/
this thing is sexy: https://openpipe.ai/?refresh=1752831956453
query rewriting :
there is problem cause now narrative engineering can happen, and also need to think about how to protect from people who is trying to gain the algo with tricks:
ML inference: https://github.com/pykeio/ort?tab=readme-ov-file
embeddings: https://github.com/Anush008/fastembed-rs
scale:
https://blog.wilsonl.in/search-engine/
jepa world model:
https://www.turingpost.com/p/jepa
text generation inference: https://github.com/huggingface/text-generation-inference
https://x.com/eugeneyan/status/1960865958991683903
large dataset: https://amazon-reviews-2023.github.io
good paper: https://arxiv.org/pdf/2305.05065
another paper: https://arxiv.org/pdf/2411.18814
https://arxiv.org/pdf/2410.06424
full course: https://decodingml.substack.com/p/33d3273e-b8e3-4d98-b160-c3d239343022
https://github.com/OpenPipe/ART
things i need to use it in the product so i can know what is working and what is not:
posthog https://www.revenuehero.io/?r=0 https://sundial.so
meter command kind of product to help user understand more about the data and get his thing done: https://www.meter.com/command
simulation based testing like tiger beetle:
https://github.com/Shopify/toxiproxy
interesting paper: https://arxiv.org/pdf/2404.03592
I love the landing page and the positioning: https://www.plain.com
competition:
shapedai
https://github.com/NVIDIA-Merlin/NVTabular https://github.com/NVIDIA-Merlin/HugeCTR?tab=readme-ov-file
end to end project:
https://github.com/jacopotagliabue/recs-at-resonable-scale
good tool we should be using for GTM: https://www.clay.com/
https://github.com/fern-api/fern
synthetic data powered simulators for beating shit out of recommendation engines:
https://arxiv.org/pdf/2309.12645 https://arxiv.org/pdf/2412.16984
cutting edge: https://arxiv.org/pdf/2509.10397v1
- User history modeling through sequence models such as Transformer
- Multi-interest user representations learning
- Multi-armed bandits model for content discovery
- Multi-tower model for learning embeddings for various kinds of entities
good lib which we can learn from on how to make lib:
https://github.com/facebookresearch/Pearl/ https://github.com/facebookresearch/ReAgent