🔧 Company Engineering Blogs

Building a GenAI Agent for Partner-Guest Messaging (booking​.ai). GenAI agent for partner-guest messaging using Python, LangGraph, FastAPI, OpenAI GPT-4 Mini, with PII guardrails and multilingual retrieval

Accelerating discovery with the AI for Math Initiative (deepmind​.google). AI for Math Initiative partners with top institutions to accelerate mathematical research using Gemini Deep Think, AlphaEvolve, and AlphaProof

ValidUntil: Ensuring compile-time data integrity in our investing system (eng​.wealthfront​.com). Wealthfront's ValidUntil framework uses Java annotations to track compile-time data validity in a component-based microservices architecture

Designing for Flow: How Leaders Create the Conditions for Team Productivity (engineering​.gusto​.com). Leadership-focused guide to designing flow: reducing cognitive load, setting rhythms, and enabling deep work for sustainable team productivity

Machine-learning predictive autoscaling for Flink (engineering​.grab​.com). Predictive autoscaling for Flink using a four-stage ML pipeline: workload forecasting, CPU usage prediction, and a custom autoscaler, applied to Kafka-backed streams at Grab

🌐 Engineering Notes

Weeknotes: 3rd November 2025 (digitalflapjack​.com). Developer diary on Yirgacheffe 1.10: CSE, edge effects, GDAL-backed raster work, Python operator overloading fixes by Michael Winston Dales

Rethinking Networking for the AI/ML Era (lukew​.com). AI/ML workloads demand high bandwidth, ultra-low latency, and near-perfect reliability, pushing networking beyond traditional models

Measuring what matters: How offline evaluation of GitHub MCP Server works (github​.blog). Offline evaluation pipeline for GitHub MCP Server evaluates tool selection, argument accuracy, and multi-tool flow readiness

Using streams of events to train machine learning models (dalelane​.co​.uk). Stream event data with Kafka and Flink to train ML models, building labeled training data from multi-stream correlations

duckdb-mlpack 0.0.3: macOS binaries, unit tests, more outputs (dirk​.eddelbuettel​.com). DuckDB-MLPack 0.0.3 adds macOS binaries, unit tests, and serialized linear regression outputs using mlpack with DuckDB

📈 Finance & Forecasting

Taming Volatility: High-Performance Forecasting of the STOXX 600 with H2O AutoML (datageeek​.com). Forecast STOXX 600 with R, tidyverse, and H2O AutoML; robust feature engineering and DRF wins

Project Spotlight: German–Hong Kong Research Collaboration on Data Analytics for Financial Markets 🇩🇪 🇨🇳 (ecmiindmath​.org). ACROSS project: data-analytics-based FBSDE solver for high-dimensional stochastic control in finance with DAAD-UGC- CUHK collaboration

I Replaced Neural Nets With XGBoost for Equity Signals. Here’s Why (medium​.datadriveninvestor​.com). Replaces neural nets with XGBoost for equity signals, highlighting calibration, NaN handling, and faster iteration

⚙️ Serving & Inference

DGX Spark and Mac Mini for Local PyTorch Development (sebastianraschka​.com). DGX Spark and Mac Mini compared for local PyTorch development, inference, benchmarking, and small-scale training with KV-cache and MPS limitations

Introduction to Serverless Model Deployment with AWS Lambda and ONNX (pyimagesearch​.com). Serverless AI deployment with AWS Lambda, API Gateway, and ONNX Runtime using Python for inference

SGLang-Jax: An Open-Source Solution for Native TPU Inference (lmsys​.org). SGLang-Jax enables native TPU inference with Jax/XLA, Ragged Paged Attention, MoE kernels, speculative decoding, and overlap scheduling

Deploying scikit-learn MLP on Tenstorrent hardware (clehaxze​.tw). Deploying scikit-learn MLP on Tenstorrent hardware using TTNN and TTMLP with Python, NumPy, PyTorch, and scikit-learn

AWS Sagemaker Bring-Your-Own-Container (BYOC) with Pixi (blog​.londogard​.com). Using Pixi in AWS SageMaker BYOC with micromamba/mamba linkage and Dockerfile tweaks

Accelerating AI inferencing with external KV Cache on Managed Lustre (cloud​.google​.com). External KV Cache on Google Cloud Managed Lustre speeds AI inference, reducing TCO and GPU needs for large-context models

🧑‍💻 Code LLM Engineering

Efficient and green LLMs for software engineering (chuniversiteit​.nl). Efficient and green LLMs for software engineering: data, model, system, and program-centric techniques for cost-effective code tasks

Composer: Building a fast frontier model with RL (simonwillison​.net). Cursor debuts Composer 1, a fast MoE model for RL-driven software engineering with tooling integration and parallel agent execution

Introducing SWE-1.5: Our Fast Agent Model (simonwillison​.net). SWE-1.5, a fast, frontier-size coding model by Windsurf with Cerebras inference at up to 950 tok/s

🖼️ Applied Vision SSL

SSL Linear Probing (ashwanirathee​.com). Self-supervised SSL linear probing with CNN encoder-decoder on masked image prediction, evaluating frozen representations across epochs

SSL based Masked Image Prediction (ashwanirathee​.com). Self-supervised masked image prediction: reconstructing full images from masked inputs using SSL concepts and related objectives

Building Production-Grade Dental AI: From Auto-Annotation to 99.5% Accuracy with YOLOv8 and NVIDIA Infrastructure (ajeetraina​.com). Production-grade dental AI with YOLOv8, NVIDIA Jetson, Docker, and auto-annotation to 99.5% mAP50 achieved

🧠 Representation & Theory

word2vec-style vector arithmetic on docs embeddings (technicalwriting​.dev). Word2vec-style vector arithmetic on full-document embeddings using EmbeddingGemma; experiments on domain transfer with customized task types and cosine similarity verification

circuit tracing (aarnphm​.xyz). Autonomous circuit tracing for causal influence in language models using transcoders, feature-level gradients, and attribution graphs

AI and the Power of Nonuniform Circuits (blog​.computationalcomplexity​.org). Discussion of AI model depths, parameter scaling, and nonuniform circuits in relation to TC0, layered architectures, and practical training trends

🧮 Math Foundations

What Is a Manifold? (quantamagazine​.org). Manifolds reshape geometry and physics, explaining how shapes look flat locally and form the backbone of relativity, topology, and data analysis

A Unit-Free Shortcut to Better Science (caltech​.edu). MIT researchers derive a dimensionless learning method IT-π to identify key variables for predictions, reducing inputs and data needs

differential foundations (aarnphm​.xyz). Bridge from topological to smooth manifolds, tangent structures, differential forms, and Morse theory with examples like TS^3, T^n, and RP^n

Notes - Uncertainty in Deep Learning MT25, Bayesian probability theory (ollybritton​.com). Bayesian probability theory notes on uncertainty in deep learning, covering rational beliefs, probability laws, and Dutch book arguments

Notes - Uncertainty in Deep Learning MT25, Introduction (ollybritton​.com). Introduction to uncertainty in deep learning, covering epistemic and aleatoric uncertainty, sources of uncertainty, and Bayesian reasoning in MT25 notes

Resampling Conserves Redundancy & Mediation (Approximately) Under the Jensen-Shannon Divergence (lesswrong​.com). Discusses resampling latent variables, redundancy and mediation under Jensen-Shannon divergence with KL comparisons and formal proofs

Flow Matching: A visual introduction (peterroelants​.github​.io). Visual introduction to Flow Matching using 1D toy example with Python (NumPy, SciPy, Matplotlib) and velocity field intuition

📚 Academic Research

Detecting Anomalies in Machine Learning Infrastructure via Hardware Telemetry (arxiv:cs). Reveal uses low-level hardware telemetry and an unsupervised pipeline to detect system and network anomalies for containerized ML workloads. Operators can optimize resources and speed up models without knowing tenant workloads

ORBIT -- Open Recommendation Benchmark for Reproducible Research with Hidden Tests (arxiv:cs). ORBIT provides standardized, reproducible recommender evaluation with public splits and a hidden large-scale ClueWeb-Reco test to stress generalization. Engineers get a transparent benchmark and leaderboard to compare real-world recommender methods

Relative Scaling Laws for LLMs (arxiv:cs). Introduces relative scaling laws tracking how performance gaps between subpopulations change with model scale using a large matched-compute Transformer sweep. Shows scaling can reduce some disparities but worsens or preserves others—important for principled scaling analyses

Panprediction: Optimal Predictions for Any Downstream Task and Loss (arxiv:cs). Panprediction formalizes training predictors that can later optimize many downstream losses/tasks and gives algorithms with provable sample complexity and calibration reductions. Crucial for ML researchers and engineers designing universal, task-agnostic predictors

A faster problem-solving tool that guarantees feasibility (news​.mit​.edu). FSNet pairs a neural predictor with a feasibility-seeking optimizer to produce constraint-satisfying solutions fast and reliably. Practical for engineers solving constrained problems (power flow, scheduling) who need guaranteed, deployable results

✨ Before you go...

You can now follow posts on the brand new ML-dedicated Mastodon and Bluesky feeds. You can also search all blog posts shown here (and MUCH more) over at https://blognerd.app.

Finally, I hope that this newsletter brings you some value. Everything here is offered free and always will be, and I'd be so grateful if you'd consider supporting me over on Patreon to help keep blaze newsletters going (if you can't afford to support financially, you can still follow there for free).

Thanks a million, and have a great week! Alastair.