Practical, technical writing on production AI: forecasting, agentic systems, RAG, MLOps, and the rest of what we work on.
Predicting the future in latent space and reconstructing it in value space score the same on a public forecasting panel. Pretraining at all is what buys accuracy under a tight labeled budget, and inside JEPA the stop-gradient does more work than the EMA target that usually gets the credit.
Read more: JEPA for time series: a labeled-budget stress test →
Reconciliation makes forecasts add up and promises better accuracy. We build the idea up from a retailer with two stores, then measure where the promise holds and where it fails on the Australian tourism hierarchy.
Read more: Forecast Reconciliation on the Australian Tourism Hierarchy: MinT, Bias and Chronos-Bolt →
TimesFM, Google's time series foundation model trained on 307B points: architecture, zero-shot inference, and benchmarks vs TiDE, Chronos, and MOIRAI.
Read more: TimesFM: The Boom of Foundation Models in Time Series Forecasting →
How Amazon's Chronos adapts language model architectures to zero-shot time series forecasting, with a hands-on benchmark against TiDE on a public dataset.
Read more: Chronos: The Rise of Foundation Models for Time Series Forecasting →36 posts
Predicting the future in latent space and reconstructing it in value space score the same on a public forecasting panel. Pretraining at all is what buys accuracy under a tight labeled budget, and inside JEPA the stop-gradient does more work than the EMA target that usually gets the credit.
Read more: JEPA for time series: a labeled-budget stress test →
Stress-testing four forecasters with RHiOTS, we find NBEATS wins on clean data but degrades four times faster under noise than AutoETS.
Read more: RHiOTS in practice: the most accurate forecaster is not the most robust →Reconciliation makes forecasts add up and promises better accuracy. We build the idea up from a retailer with two stores, then measure where the promise holds and where it fails on the Australian tourism hierarchy.
Read more: Forecast Reconciliation on the Australian Tourism Hierarchy: MinT, Bias and Chronos-Bolt →
How Combining Foundational and Non-Foundational Models Can Take Forecasting Further
Read more: FORGE: A Hybrid Framework Bridging Foundational and Non-Foundational Models for Forecasting →
How BLIP-2 works and how to combine it with Gemini to build a multimodal fashion search agent that finds products from text and image queries.
Read more: Multimodal Search Engine Agents Powered by BLIP-2 and Gemini →
Step-by-step guide to building visual agents that browse the web autonomously, using multimodal LLMs to reach goals and retrieve information without APIs.
Read more: Building Visual Agents that can Navigate the Web Autonomously →
A step-by-step guide to building a multi-agent framework from scratch, covering planning, tool use, and orchestration without leaning on off-the-shelf agent libraries.
Read more: Agentic AI: Building Autonomous Systems from Scratch →TimesFM, Google's time series foundation model trained on 307B points: architecture, zero-shot inference, and benchmarks vs TiDE, Chronos, and MOIRAI.
Read more: TimesFM: The Boom of Foundation Models in Time Series Forecasting →
How Llama's architecture evolved from Llama 1 to 3.1 — GQA, tokenizer, and context changes — plus a hands-on Llama 2 vs Llama 3 comparison using RAQ.
Read more: The Evolution of Llama: From Llama 1 to Llama 3.1 →
RAQ, a statistical framework for ranking LLM answer quality, plus a guide to deploying open-source models at scale with NVIDIA NIM microservices.
Read more: Research to Production: Relative Answer Quality (RAQ) and NVIDIA NIM →
Build a Bayesian marketing mix model in PyMC to measure channel performance, estimate ROAS, and optimize budget allocation with adstock and saturation.
Read more: MMM: Bayesian Framework for Marketing Mix Modeling and ROAS →
Inside Moirai, Salesforce's universal time series foundation model: its architecture, zero-shot forecasting, and a benchmark against TiDE and Chronos.
Read more: Moirai: Time Series Foundation Models for Universal Forecasting →How Amazon's Chronos adapts language model architectures to zero-shot time series forecasting, with a hands-on benchmark against TiDE on a public dataset.
Read more: Chronos: The Rise of Foundation Models for Time Series Forecasting →
TimeGPT vs TiDE: Is Zero-Shot Inference the Future of Forecasting or Just Hype?
Read more: TimeGPT vs TiDE →
How FinalMLP's two-stream MLP with feature gating and bilinear fusion improves CTR prediction, plus a hands-on book recommendation implementation.
Read more: FinalMLP: A Simple yet Powerful Two-Stream MLP Model for Recommendation Systems →
A walkthrough of Meta's Seamless translation family — SeamlessM4T v2, Expressive, and Streaming — with architectures and a Python guide using HuggingFace.
Read more: Seamless: In-Depth Walkthrough of Meta’s New Open-Source Suite of Translation Models →
TFT: an Interpretable Transformer. Graphic representation of TFT, an interpretable transformer, showcasing its architecture and features.
Read more: TFT: an Interpretable Transformer →
Mistral AI vs. Meta: Comparing Top Open-source LLMsA comparison between Mistral 7B vs Llama 2 7B and Mixtral 8x7B vs Llama 2 70B
Read more: Mistral AI vs. Meta: Comparing Top Open-source LLMs →
Exploring the Transformer’s Decoder Architecture: Masked Multi-Head Attention, Encoder-Decoder Attention, and Practical Implementation
Read more: LLMs and Transformers from Scratch: the Decoder →
AI-Powered Customer Support App: Semantic Search with PGVector, Llama2 with an RAG System, and Advanced Translation Models
Read more: AI-Powered Customer Support App →
Stable Diffusion: Mastering the Art of Interior Design: A deep dive into Stable Diffusion and its inpainting variant for interior design
Read more: Stable Diffusion: Mastering the Art of Interior Design →A deep exploration of TiDE, its implementation using Darts and a real life use case comparison with DeepAR and TFT
Read more: TiDE: the ‘embarrassingly’ simple MLP that beats Transformers →
Mastering Multilingual Audio Transcription and Semantic Search: Unlock the potential of cross-language information accessibility
Read more: Mastering Multilingual Audio Transcription and Semantic Search →Exploring the intricacies of encoder, multi-head attention, and positional encoding in large language models
Read more: Creating Transformer Encoders and Multi-Head Attention Layers in Python from Scratch →
A deep dive into tailoring pre-trained LLMs for custom use cases using a RAG approach, featuring LangChain and Hugging Face integration
Read more: Retrieval Augmented Generation →
GPT-4V and Text-to-Speech Technology Enhancing Visual Impairment Navigation: Integrating GPT-4V and TTS for Advanced Sensory Assistance
Read more: Visually Impaired with GPT-4V and Text-to-Speech Technology →
Llama2 for Python Coding Enhancing Llama2’s proficiency in Python through supervised fine-tuning and low-rank adaptation techniques
Read more: Fine-Tune Llama2 for Python Coding on Consumer Hardware →
Open vs Closed Source LLMs, Unpacking Llama 2’s Unique Features, Mastering the Art of Prompt Engineering, and Designing Robust Solutions
Read more: Leveraging Llama 2 Features in Real-world Applications →
Document-Oriented Agents: Leveraging ChromaDB, Langchain, and ChatGPT: Enhanced Responses and Cited Sources from Large Document Databases
Read more: Document-Oriented Agents →
OpenAI’s Function Calling in Language Learning Models: Transforming Data Pipelines with OpenAI’s Function Calling Feature: Implementing an Email Sending Workflow Using PostgreSQL and FastAPI
Read more: OpenAI’s Function Calling in Language Learning Models →
Harnessing the Falcon 40B Model, the Most Powerful Open-Source LLM: Mastering open-source language models: diving into Falcon-40B
Read more: Falcon 40B Model, the Most Powerful Open-Source LLM →
Speech Model that Supports 1162 Languages: Explore the cutting-edge multilingual features of Meta’s latest automatic speech recognition model
Read more: Testing the Massively Multilingual Speech Model →
Vosk for Efficient Enterprise-Grade Speech Recognition: A comprehensive walkthrough of implementing and assessing Vosk-based speech recognition systems
Read more: Vosk for Efficient Enterprise-Grade Speech Recognition →
Whisper JAX vs PyTorch: Uncovering the Truth about ASR Performance on GPUs
Read more: Uncovering the Truth about ASR Performance on GPUs →
Unlock the Power of Audio Data: Advanced Transcription and Diarization with Whisper, WhisperX, and PyAnnotate
Read more: Unlock the Power of Audio Data →
Master Semantic Search at Scale: Index Millions of Documents with Lightning-Fast Inference Times using FAISS and Sentence Transformers
Read more: Master Semantic Search at Scale →Tell us about the decision you're trying to make. We'll tell you whether Horizon, Pliney, or a Custom AI Solution fits, or whether it's none of the above.