Apertus – Open Foundation Model for Sovereign AI

Apertus is a fully open-source LLM initiative from EPFL, ETH Zurich, and CSCS, providing open weights, data, and documented training methods. Available in 8B and 70B parameter scales, the models are multilingual across 1000+ languages and engineered for EU AI Act compliance via PII removal and memorization prevention.

Read more

When I reject AI code even if it works

The primary bottleneck in AI-assisted development has shifted from implementation to the cognitive load of reviewing code generated without the developer's initial mental mapping. Engineers frequently reject LLM outputs that lack explainability, introduce premature abstractions, or increase system complexity despite passing CI. Human oversight remains essential to guide coding agents toward sustainable, scalable architectures that prioritize long-term maintainability over immediate implementation speed.

Read more

Building reliable agentic AI systems

Bayer’s PRINCE is an agentic RAG platform designed to streamline preclinical drug discovery by transitioning from keyword search to a multi-agent "ask and do" system. Orchestrated via LangGraph, the architecture utilizes specialized agents for research (combining hybrid OpenSearch retrieval and Text-to-SQL), data reflection, and synthesis. The system emphasizes context and harness engineering, implementing state persistence, LLM fallbacks, and RAGAS-based evaluation to ensure reliability and traceability in a regulated environment.

Read more

The 100k whys of AI

LLMs are distinguishable from human authors not by individual linguistic errors, but through the quasi-deterministic repetition of specific patterns across similar prompts. This convergence results in "AI slop," characterized by functionally identical outputs like the clusters of near-identical book titles and covers found on Amazon. While individual samples may seem human, the collective predictability of these models provides a fuzzy but recognizable signal of AI generation.

Read more

The early hiring funnel is now breaking on both ends

Generative AI has compromised traditional hiring funnels by enabling automated resume optimization and real-time interview assistance, with data showing suspicious patterns in up to 60% of entry-level technical screenings. This shift renders standard behavioral interviews ineffective, as candidates can use LLM-powered overlays to script responses during live video calls. To maintain talent density, organizations must transition to adaptive, reasoning-based assessments and "AI-enabled" interviews that prioritize human judgment and the ability to critically evaluate LLM outputs over manufactured polish.

Read more

The largest open database of local laws in the US

LOCUS is a comprehensive, machine-readable corpus of 9,239 U.S. municipal and county ordinance codes designed to facilitate legal AI research. The dataset uses OCR to standardize fragmented local laws and includes a county-harmonized layer covering the majority of the U.S. population. Alongside the corpus, the authors released ModernBERT-based classifiers to analyze legal dimensions such as opacity and paternalism at scale.

ArXiv

Refloow Photo Studio – A local, offline photo editor with on-device AI

Refloow Photo Studio is a free, open-source desktop photo editor built with Node.js and Electron, designed for accessible, professional-looking image manipulation. It emphasizes privacy by running entirely locally without accounts or data collection, offering features like instant background removal, color adjustments, layering, and over 40 filters. The project aims to simplify complex editing workflows while preserving original image quality.

Repo

Daily_stock_analysis: LLM-powered multi-market stock analysis system

This LLM-powered intelligent stock analysis system provides automated daily analysis and "decision dashboards" for A-shares, HK, US, Japan, and Korea markets. It integrates various LLMs (e.g., OpenAI compatible, Gemini, Claude, Ollama) and diverse data sources, including market data, news, and social sentiment. Key features include AI decision reports, multi-market data aggregation, an Agent-based multi-turn strategy questioning system, and a web UI for manual analysis, backtesting, and portfolio management, with flexible deployment options via GitHub Actions or Docker.

Repo

DebugBrief – turn debugging sessions into reports, no AI

DebugBrief is a CLI tool that records debugging workflows to generate evidence-backed Markdown reports for PRs, handoffs, and incident notes. It captures command outputs, exit codes, and Git metadata to document "red-to-green" transitions and correlated file changes. The tool operates deterministically without AI, ensuring reports are based solely on recorded terminal activity and repository state.

Repo

OpenMontage: Turn your AI coding assistant into a full video production studio

OpenMontage is an open-source, agentic video production system that enables AI coding assistants to orchestrate the entire lifecycle of video creation, from research and scripting to asset generation and final composition. It features an agent-first architecture using Python tools and Markdown-based skills to manage 12 specialized pipelines, including animated explainers and documentary montages. The system integrates production-grade quality gates, budget governance, and a multi-provider scoring engine that supports both cloud APIs and local GPU workflows.

Repo

Taste turns your Git history into enforceable coding judgment

taste is a CLI tool that compresses AI agent context by up to 97% by extracting project-specific patterns from git history and session logs. It automatically learns naming conventions, architectural styles, and "banned patterns," persisting them in TASTE.md and .agent-taste.json to ensure LLM consistency. By packing essential context into a compact .session-doc.md, it reduces token overhead while improving code generation quality across various agent harnesses.

Repo