Machine learning engineer and AI scientist with a Master's in theoretical physics, based in Los Angeles. Experienced in developing production ML systems and agentic applications, with a track record of technical leadership and client collaboration.
Writing on RAG, LLM systems, and whatever else I'm building or thinking about.
-
Should Your Intent Router Be a Decision Model?
I built a decision model from a small open LLM, then raced it against an embedding model, an embedding classifier and Claude Haiku 5.5 as an intent router. The answer, plus the setup details that move a DIY decision model by 20 to 30 points.
-
Exploring EmbeddingGemma 2
A first poke at Google's EmbeddingGemma 2: whether task prefixes matter, why keyboard mash can score closer to a query than a real sentence does, how far you can truncate the vectors, and how it handles topics and languages.
-
Citation Prompting: Integrating Retrieval Results with LLMs in RAG
How to structure retrieved context in the prompt, and how to evaluate whether the generation step is actually using it well.
-
Choosing Embedding Models for RAG
Why the default embedding model is often the wrong one, and how to evaluate candidates for your domain.
-
You Should Probably Be Using Hybrid Search
Semantic search alone has blind spots. Why combining it with keyword search consistently wins.
-
Why Chunking Matters for Effective RAG
Fixed-size chunking is a common pitfall. A look at recursive, sentence, semantic, and structure-aware alternatives.
-
What do larger context windows really mean for RAG?
Long-context LLMs don't obviate RAG. Why retrieval stays essential as context windows grow.
-
RAG in Production: Avoiding Common Pitfalls
Common failure modes when moving RAG from prototype to production, and how to mitigate them.
Papers
-
Narrative Embedding: Re-Contextualization Through Attention
Narrative analysis is becoming increasingly important for a number of linguistic tasks including summarization, knowledge extraction, and question answering. We present a novel approach for narrative event representation using attention to re-contextualize events across the whole story. Comparing to previous analysis we find an unexpected attachment of event semantics to predicate tokens within a popular transformer model. We test the utility of our approach on narrative completion prediction, achieving state of the art performance on Multiple Choice Narrative Cloze and scoring competitively on the Story Cloze Task.
-
Sequence or Pseudo-Sequence? An Analysis of Sequential Recommendation Datasets
Sequential recommendation aims to model a user's preferences by looking at the order of interactions in a user's history. The evaluation of such algorithms requires robust datasets with genuine sequential information. In this work we analyze the timestamp information of several commonly used datasets and show that reported timestamps are not indicative of meaningful sequential order. In the datasets explored, significant numbers of users have interactions occurring at identical timestamps. The actual order of these interactions is therefore unknowable; the interaction history is pseudo-sequential. We find that randomly shuffling the order of interactions has minimal impact on the performance of a leading sequential recommender. Particular attention is paid to MovieLens because of its frequency of use in the field of sequential recommendation. Our findings motivate the necessity for new datasets with more meaningful ordering for the evaluation of sequential recommenders.
-
Anamorphic Quasiperiodic Universes in Modified and Einstein Gravity with Loop Quantum Gravity Corrections
The goal of this work is to elaborate on new geometric methods of constructing exact and parametric quasiperiodic solutions for anamorphic cosmology models in modified gravity theories, MGTs, and general relativity, GR. There exist previously studied generic off-diagonal and diagonalizable cosmological metrics encoding gravitational and matter fields with quasicrystal like structures, QC, and holonomy corrections from loop quantum gravity, LQG. We apply the anholonomic frame deformation method, AFDM, in order to decouple the (modified) gravitational and matter field equations in general form. This allows us to find integral varieties of cosmological solutions determined by generating functions, effective sources, integration functions and constants. The coefficients of metrics and connections for such cosmological configurations depend, in general, on all spacetime coordinates and can be chosen to generate observable (quasi)-periodic/ aperiodic/ fractal / stochastic / (super) cluster / filament / polymer like (continuous, stochastic, fractal and/or discrete structures) in MGTs and/or GR. In this work, we study new classes of solutions for anamorphic cosmology with LQG holonomy corrections. Such solutions are characterized by nonlinear symmetries of generating functions for generic off-diagonal cosmological metrics and generalized connections, with possible nonholonomic constraints to Levi-Civita configurations and diagonalizable metrics depending only on a time like coordinate. We argue that anamorphic quasiperiodic cosmological models integrate the concept of quantum discrete spacetime, with certain gravitational QC-like vacuum and nonvacuum structures. And, that of a contracting universe that homogenizes, isotropizes and flattens without introducing initial conditions or multiverse problems.
Check out my band and our music below.
Feel free to reach out to me at danieljwoolridge [at] gmail [dot] com