Blog

Posts on generative models, LLM systems, applied machine learning and ML for science.
Date Title Categories Description
Oct 9, 2026 AI agents on the KLS conjecture: a search mathematicians can trust and join AI for mathematics, agents AI agents can now produce more mathematics than anyone can check. Human–AI Mathematics is an open framework with two aims: a structure that keeps an agent-driven search traceable, where a claim moves to proved only through an independent review of its exact text, and a manuscript that mathematicians can read, check and contribute to. Its first case study is the KLS conjecture, which three teams proved while the search was running.
Aug 26, 2026 Mills ratios on manifolds: what Gaussian smoothing sees near a boundary asymptotics, diffusion models Convolving a singular measure with a Gaussian of width σ produces a smooth density whose logarithmic derivatives can blow up as σ → 0. Zooming in at scale σ replaces the support by its inward tangent cone: its Gaussian mass gives the leading density, and derivatives of its logarithm give the leading singular fields — a manifold-valued inverse Mills ratio.
Jul 20, 2026 Closed-form flow matching: posterior concentration and softmax collapse deep learning, flow matching The closed-form velocity field of empirical flow matching is a softmax over training examples. Reading that softmax as a Bayesian posterior turns its collapse into a decoding problem, with an information-theoretic threshold that can be evaluated on CIFAR-10.
May 12, 2026 Flow matching as posterior averaging deep learning, flow matching A velocity-based introduction to flow matching with Gaussian probability paths: the marginal velocity field is a posterior average of conditional targets, and the gap between the conditional and marginal losses is an irreducible posterior-variance term that can be computed in closed form.
Jan 18, 2026 Analyzing Raman spectra with machine learning machine learning, chemistry From inelastic scattering to a calibrated regression model: the physics behind a Raman spectrum, the preprocessing pipeline that makes real spectra usable, and a PLS regression baseline for predicting analyte concentrations.
Aug 15, 2025 Generating slide decks with an LLM: a comparison of formats LLM One real presentation, four formats, one model: what Gemini 2.5 Pro actually produces when asked to rebuild an EDF producer booklet in HTML/CSS, Quarto/Reveal.js, the Google Slides API, and python-pptx — and why the failures in the API-driven formats are silent ones.
Jun 30, 2025 LLM observability with Langfuse and a FastHTML chatbot LLM, python Instrumenting a chatbot end to end: Langfuse traces a conversation as a span, each model turn as a generation, and each thumbs-up as a score — wired into a FastHTML app that updates over a WebSocket instead of rerunning a script.
Mar 20, 2025 PDF RAG: recovering document structure and querying a corpus document AI, LLM, RAG Page-by-page parsing destroys the structure a reader relies on. This post rebuilds it for PDFs with a vision language model, extracts metadata and a document tree, and then queries a corpus of reports and slide decks with a two-level ReAct agent.
Feb 18, 2025 Parsing PDFs for LLM input document AI, LLM How Docling, Marker, MinerU, pymupdf4llm, LlamaParse, Gemini, and Mistral OCR turn PDFs into Markdown: what the pipelines are made of, and where each one silently gets it wrong on six real documents.
Feb 6, 2025 Predicting MHC-peptide presentation with machine learning deep learning, biology Training a neural network to predict which peptides a given MHC class I allele presents, and benchmarking it against MHCflurry.
Apr 24, 2024 Encoding distances in molecules and pockets: GBFPT versus DCEPT deep learning, biology Uni-Mol encodes interatomic distances with a Gaussian basis whose parameters receive very large gradients. Replacing it with a discretized categorical embedding, in the spirit of AlphaFold, stabilizes pretraining but degrades the pocket embeddings used for retrieval.
No matching items