Built a RAG assistant that does it for you - ask a question, get a cited answer pulled from your own PDF library. This is a hybrid retrieval pipeline for scientific papers:
> Dense: cosine similarity over OpenAI embeddings in pgvector → top 20
> Sparse: keyword ILIKE → top 20
> RRF fusion merges both rankings
> Similarity threshold filter (with fallback)
> LLM reranker scores each chunk 0–10
> Generator answers with inline citations -or says 'unsupported' if evidence is weak
Live demo: https://t.co/d6gfRxksbv
GitHub: https://t.co/2wWjC7RMdA
@trq212 Looking forward to your article! It's intriguing how Opus 4.8 combines intelligence with collaboration. Would love to hear more about these workflows.
@AnthropicAI Chris Olah at the presentation of an encyclical is significant! Curious to hear his insights on the intersection of human values and AI ethics discussed there.
@arankomatsuzaki Interesting insight on parallel agents! It's impressive to see the challenges of sequential problems laid out like that almost like trying to untangle a single thread from a tightly wound ball.
VAE model day:
- 1D CNN compresses 30-cycle window to latent vector (dim=16)
- Decoder's MSE shows anomaly; trained on healthy data, degraded struggle
- Delta features = major breakthrough
- Precision=0.38, Recall=0.88, F1=0.53, AUROC=0.75
- A unique trade-off: broad net with some false alarms
When you are not able to concentrate but you have to start working on your next project:
A real time anomaly detection system for industrial sensor streams.
Multimodal:
- Numerical time series (vibrations, temperature, pressure, current)
- Audio (mel-spectograms from machine sounds)
- text logs
Reading 20 papers to answer one research question is brutal.
Built a RAG assistant that does it for you - ask a question, get a cited answer pulled from your own PDF library. This is a hybrid retrieval pipeline for scientific papers:
> Dense: cosine similarity over OpenAI embeddings in pgvector → top 20
> Sparse: keyword ILIKE → top 20
> RRF fusion merges both rankings
> Similarity threshold filter (with fallback)
> LLM reranker scores each chunk 0–10
> Generator answers with inline citations -or says 'unsupported' if evidence is weak
FastAPI + Next.js + Postgres + Redis. Fully open source.
https://t.co/ohU6QMwtfR