📰 Medium · Data Science

Proprietary Deep research agents just became much more available! No need for perplexity or Claude code anymore, build your own! Continue reading on Towards AI

1. What is synthetic data in the context of generative AI? Continue reading on Medium »

Position, verbosity, self-preference, format, and calibration drift are separate failure modes with separate mechanisms. The mitigations… Continue reading on IL

A practical guide to reranking in Retrieval-Augmented Generation, why bi-encoders alone are not enough, and how to use cross-encoders in… Continue reading on Me

A world model and a large language model face the same question. We follow both, step by step, by hand. Then we ask what “thinking” would… Continue reading on D

A step-by-step breakdown of how multilingual embeddings work. Continue reading on Medium »

Welcome to Day 1 of 60. Continue reading on Medium »

ModelPulseX watches 29 free models from OpenCode Zen and OpenRouter every five minutes and tells you which one is actually fast right now. Continue reading on M

Every⁠one knows tha⁠t reword​ing a pro‌mpt can transform an AI’s answer,‍ and the usual explanation is that the mode​l is “sensitive‍”… Continue reading on Towa

Building a fact-checking pipeline with web search, open-source tools, Python, and AI agents Continue reading on Medium »

Your LLM passed the benchmark. That doesn’t mean your application works. Continue reading on CodeToDeploy »

The AI ecosystem is no longer just about typing prompts into a web browser. Behind the scenes, a massive engineering transformation is… Continue reading on Medi

How semantic vector retrieval surfaces deprecated payer guidelines, and how to build deterministic temporal version proxies and dependency… Continue reading on

Kimi K3 is a 2.8T parameter Mixture-of-Experts model with 104 billion activated parameters, native vision capabilities, and a… Continue reading on Medium »

When I first started learning about Agentic AI, I noticed that terms like LLM, chatbot, workflow, and AI agent were often used so loosely… Continue reading on M

How an unheralded, free 1M-token multimodal reasoning model surfaced on OpenRouter and OpenCode to disrupt the frontier AI landscape. Continue reading on Toward

How Anthropic’s open standard solved the LLM tool-calling fragmentation problem and revolutionized developer tooling. Continue reading on Medium »

A simulation of teacher-student distillation project. And LLM as a judge failure modes, I debuged along the way. Continue reading on Medium »

A high schooler’s attempt to measure whether ChatGPT-style agreement predicts market overreaction, and why the honest answer is “maybe… Continue reading on Medi

After 3 years in the trenches, working under US university professors, building for multiple AI startups, and sitting across the table… Continue reading on Medi

A base M4 Mac mini with 16GB runs useful language models locally and serves them across a network. Here’s the full setup, the speeds it… Continue reading on Med

When a sentence misleads a reader, the eyes give it away: they stop and jump back to reread an earlier word. A new study led by William… Continue reading on Med

Retrieval-augmented generation lets a model look things up instead of trying to remember everything. Continue reading on Medium »

Imagine an organisation approves a model. Continue reading on Medium »