# Divinci AI — full content index

> Expanded companion to https://divinci.ai/llms.txt. Every English page in the
> site search index, with its summary. Translations (es, fr, ar, de, it, pt,
> ru, ja, zh, ko, nl, hi) mirror this structure under their language prefix and
> are listed in https://divinci.ai/sitemap.xml.


## Artificial Intelligence

- [The Future of RAG Systems: Beyond Simple Document Retrieval](https://divinci.ai/blog/future-of-rag-systems/): Where RAG is heading: scored-QA routing, vector arenas, live competitive evaluation. May 2026 — RAG is now a routing problem, not a pipeline one. Retrieval-Augmented Generation (RAG) has emerged as one of the most transformative applications of Large Language Models (LLMs), enabling AI systems to ac

## Company News

- [Divinci AI Joins Cloudflare Workers Launchpad Cohort #6](https://divinci.ai/blog/cloudflare-workers-launchpad-cohort-6/): Divinci AI joined Cloudflare's Workers Launchpad Cohort #6. Sub-100ms edge-RAG, the Demo Day pitch, and a deep-dive on our production stack. Building AI at the speed of light on Cloudflare's global edge network For over 15 years, we've trusted Cloudflare. Their ever-free tier grants you the world's

## Compliance

- [Validating and Releasing Custom LMs in Regulated Fields](https://divinci.ai/blog/validating-and-releasing-custom-lms-in-regulated-fields/): EU AI Act, GDPR Article 17, HIPAA, NIST AI RMF — mapped capability-by-capability to an LLM release pipeline. Where open vs closed weights diverge. Notes from the Release Cycle — Part IV --- A general counsel walks into the engineering review. She has one question: "If the EU AI Act Article 17 right-

## Engineering

- [Hosted Hermes on Cloudflare: One Agent, One Sandbox](https://divinci.ai/blog/hosted-hermes-on-cloudflare/): We now host NousResearch Hermes agents inside Divinci — one isolated Cloudflare Sandbox container per agent, chattable in-app or connectable from a local Hermes over an OpenAI-compatible proxy. Here's what it is, who it's for, and how the isolation and security actually work. The short version You c
- [Putting a Grounded Assistant on a Real Phone Number](https://divinci.ai/blog/grounded-voice-agents-real-phone-calls/): We put our grounded assistant on a real phone number, then cut first-audio latency from ~7s to ~4.5s — by measuring the call path, not guessing at it. --- TL;DR We put a RAG-grounded, fine-tuned, medically-guardrailed assistant on a real phone number. Calls ride a Twilio SIP trunk into a LiveKit roo
- [We Open-Sourced the Pipeline That Builds Our Demos](https://divinci.ai/blog/open-sourcing-the-demo-pipeline/): The pipeline that researches a company, crawls its site, builds a white-label RAG demo, and delivers a working link is now Apache-2.0 on GitHub — gates, crawl policy, agent skills and all. When we want to show a company what Divinci would do with their content, we don't mock it up. We build it. A pi

## Other

- [AI Compliance — Powered by vIndex](https://divinci.ai/compliance/): Verifiable AI compliance for the EU AI Act, GDPR Article 17, HIPAA, and NIST AI RMF. The vIndex is the technical-documentation artifact regulators are about to require — and Divinci ships with it.
- [AI Safety, Trust and Ethics Statement](https://divinci.ai/ai-safety/): Divinci AI's Commitment to Responsible and Trustworthy AI At Divinci AI, we prioritize safe, ethical, and transparent AI solutions. Our products, including web and mobile applications, serve diverse use cases in healthcare and other sensitive fields. This document details our commitment to using onl
- [AI Voice Agents — Your Grounded Assistant, On a Real Phone Number](https://divinci.ai/voice-agents/): Give your Divinci release a phone number. Callers get the same RAG-grounded, guardrailed, fine-tuned assistant you ship on the web — streaming answers in seconds over an ordinary phone call. No app, no browser, no sign-up. / Page-specific Leonardo journal background — reuses the AutoRAG art / .featu
- [API Reference](https://divinci.ai/api/): Complete REST API reference for Divinci AI — 60+ endpoints for managing releases, RAG knowledge bases, fine-tuning, transcripts, and more. / Full-width Redoc container — no page chrome / .feature-page.leonardo-bg::before { display: none !important; } redoc-container { margin: 0 -2rem; min-height: 80
- [About Us](https://divinci.ai/about/): Divinci AI builds the safety, security and governance layer for custom language models and agents. Meet the team and the mission. This content will be rendered by the about.html template.
- [Accessibility Statement](https://divinci.ai/accessibility/): Divinci AI's commitment to digital accessibility, WCAG 2.1 conformance targets, the standards we follow, and how to report barriers. Accessibility Statement Divinci AI is committed to ensuring digital accessibility for people with disabilities. We are continually improving the user experience for ev
- [AutoRAG - Automated Retrieval Augmented Generation](https://divinci.ai/autorag/): Automatically find the optimal RAG pipeline for your data with Divinci AI's comprehensive AutoRAG solution / Page-specific Leonardo journal background / .feature-page.leonardo-bg::before { background-image: url('/images/bg-autorag.svg') !important; background-repeat: no-repeat !important; background
- [Brand](https://divinci.ai/brand/): Divinci AI brand assets, logo downloads, color palette, and usage guidelines.
- [CLI Reference](https://divinci.ai/cli/): Complete command-line reference for the Divinci AI CLI — 55+ commands auto-generated from source. Manage workspaces, chat, knowledge bases, releases, and more from your terminal.
- [Careers](https://divinci.ai/careers/): Join the Divinci AI team and help shape the future of human-AI collaboration. Explore career opportunities across engineering, product, design, and business roles. / Careers page specific styles / .careers-hero { background: linear-gradient(135deg, rgb(248, 244, 240) 0%, rgb(240, 235, 225) 100%); co
- [Changelog](https://divinci.ai/changelog/): Divinci AI's changelog tracks our product updates, new features, and improvements over time to keep our users informed about our development progress. / Page-specific Leonardo journal background / .leonardo-bg::before { background-image: url('/images/bg-changelog.svg') !important; background-repeat:
- [Contact](https://divinci.ai/contact/): Get in touch with Divinci AI. Contact our team for questions about our platform, pricing, or to request a demo. This content will be rendered by the contact.html template. EOF < /dev/null
- [Cookie Policy](https://divinci.ai/cookies/): Cookie usage and privacy policy for Divinci AI website Cookie Policy What are cookies? Cookies are small text files that are placed on your computer or mobile device when you visit a website. They are widely used to make websites work more efficiently and provide information to website owners. How w
- [Data Processing Agreement](https://divinci.ai/data-processing-agreement/): Data Processing Agreement (DPA) between Divinci AI and businesses using Divinci to power a whitelabel or embedded AI assistant for their own end users. Who this is for: This DPA applies to businesses ("Customers") that use Divinci to power an AI assistant under their own brand — an embedded chat wid
- [Developer Tools & Documentation](https://divinci.ai/docs/): Comprehensive developer documentation for Divinci AI — CLI, Server SDK, Client SDK, MCP SDK, Embed Client, and REST API reference. / Page-specific Leonardo journal background / .feature-page.leonardo-bg::before { background-image: url('/images/bg-api.svg') !important; background-repeat: no-repeat !i
- [Divinci Local Inference — Privacy Policy](https://divinci.ai/local-inference-privacy/): Privacy policy for the Divinci Local Inference Chrome extension: what runs locally on your device and what, in specific signed-in situations, is sent to Divinci. Divinci Local Inference — Privacy Policy Last updated: June 2026 This policy applies specifically to the Divinci Local Inference Chrome ex
- [Divinci Local Inference — run AI in your browser, on your own GPU](https://divinci.ai/local-inference/): A Chrome extension that runs open-weight models like Gemma 4 and Llama 3.2 entirely on your machine via WebGPU. Private by default, no API cost, works offline once loaded. / Cycling model name in the hero. Pure CSS — no inline script, works with JS disabled, and degrades to the first item under redu
- [Governed AI Release Management - Approvals, Versioning, Rollback](https://divinci.ai/release-management/): Governed release management for AI models and agents: human sign-off, versioned releases, instant rollback, and an audit trail of who approved what and when. / Page-specific Leonardo journal background / .feature-page.leonardo-bg::before { background-image: url('/images/bg-release.svg') !important;
- [Hosted Hermes Agents — Your Own Hermes, Isolated in the Cloud](https://divinci.ai/hermes-agents/): Run your own NousResearch Hermes agent inside Divinci: an isolated Cloudflare Sandbox container per agent, one Durable Object each. Chat in-app, or connect a local Hermes, the desktop app, or any OpenAI-compatible client through a per-agent proxy URL. Bring your own provider key. / Reuse the Leonard
- [LLM Safety and Quality Assurance - Evaluate Before Release](https://divinci.ai/quality-assurance/): Safety and quality evaluation for custom LLMs and agents: automated regression, red-team and domain test suites, calibrated judges, and continuous monitoring, with the evidence retained for auditors. / Page-specific Leonardo journal background / .feature-page.leonardo-bg::before { background-image:
- [Press](https://divinci.ai/press/): Press resources for Divinci AI including news, media coverage, press releases, brand assets, and contact information for press inquiries. This content will be rendered by the press.html template. EOF < /dev/null
- [Pricing Plans](https://divinci.ai/pricing/): Divinci AI offers flexible pricing plans for businesses of all sizes. Choose from our Starter, Pro, and Enterprise plans to create custom AI solutions for your organization. / Page-specific Leonardo journal background / .feature-page.leonardo-bg::before { background-image: url('/images/bg-pricing.sv
- [Privacy Policy](https://divinci.ai/privacy-policy/): Divinci AI's commitment to protecting your privacy and personal data in compliance with GDPR and international privacy laws Last updated: July 2026 Our Commitment to Privacy At Divinci AI, we are committed to protecting your privacy and ensuring the security of your personal data. This Privacy Polic
- [Product Roadmap](https://divinci.ai/roadmap/): Divinci AI's product roadmap details our upcoming features, enhancements, and strategic direction for our AI collaboration platform. / Page-specific Leonardo journal background / .leonardo-bg::before { background-image: url('/images/bg-roadmap.svg') !important; background-repeat: no-repeat !importan
- [RAG Arena & Dynamic Routing](https://divinci.ai/rag-arena/): Compare knowledge bases side-by-side and automatically route questions to the best-performing RAG vector with Divinci AI's intelligent arena system / RAG Arena Page Styles - Renaissance Warm Palette / / Page-specific Leonardo journal background / .feature-page.leonardo-bg::before { background-image:
- [RAG Routing — One API, Many Architectures](https://divinci.ai/rag-routing/): Divinci's RAG Routing dispatches every query to the cheapest backend that can answer it correctly. Ten supported retrieval engines (PageIndex, neo4j-hybrid, RAPTOR, LightRAG, Qdrant, Cloudflare Vectorize, Couchbase, Vertex AI, MongoDB Atlas, Redis Vector) behind one endpoint, with learned per-questi
- [Security](https://divinci.ai/security/): How Divinci AI protects your data — de-identification, access control, audit logging, and honest answers about where we are on formal certifications. Security is core to how we build. This page describes what's actually true about our architecture and practices today — not a marketing checklist. Whe
- [Squarespace → Cloudflare + AI — Convert your site with Divinci](https://divinci.ai/squarespace-to-cloudflare/): Turn your Squarespace site into a fast, self-hosted Cloudflare Workers site with a built-in AI chat assistant. Divinci's Site Converter screenshots your live page and regenerates clean, editable HTML with a vision model — then you own the code. .section-padding { padding: 4rem 0; } .section-heading
- [Support Center](https://divinci.ai/support/): Get help and support for Divinci AI. Find answers to common questions, access tutorials, and connect with our support team. This content will be rendered by the support.html template.
- [System Status](https://divinci.ai/status/): Live availability status for the Divinci AI platform, derived from our infrastructure monitors.
- [Terms of Service](https://divinci.ai/terms-of-service/): Terms and conditions for using Divinci AI services Important: Please review our updated Terms of Service carefully. 1. Introduction Welcome to Divinci AI, Inc. (" Company ", " we ", " us ", or " our "). We provide an AI platform that includes web, desktop, and mobile applications, embeddable AI chat
- [The Open Web Vector Initiative](https://divinci.ai/open-web-vectors/): A public, per-site retrieval index for the open web: every site gets its own vector database, its own embeddings, and a chat endpoint grounded in its own words — with the permission check running before the crawler, not after.
- [The RAG Universe](https://divinci.ai/www-rag/): Browse Divinci's curated catalog of web sources, each indexed and ready to chat with — no sign-up required. Or copy any index into your own workspace and build an AI assistant on top of it.
- [Tutorials](https://divinci.ai/tutorials/): Learn how to get the most out of Divinci AI with our comprehensive tutorials Tutorials Welcome to our comprehensive tutorial library. Here you'll find step-by-step guides to help you maximize your use of Divinci AI's tools and services. Getting Started Quick Start Guide Learn the basics of setting u
- [Voice Agent Call Transcripts — Real Test-Line Dialogues](https://divinci.ai/voice-agent-scripts/): Lightly-trimmed transcripts from real calls to a Divinci Voice Agent on our staging test line — showing grounded answers, multi-turn memory, and medical-safety deferral over an ordinary phone call. Voice Agent call transcripts These are lightly-trimmed transcripts from real calls and turn-by-turn te

## Policy

- [Universal Basic Income by 2035: A Feasible Path Forward](https://divinci.ai/blog/universal-basic-income-2035/): UBI by 2035: how AI productivity, energy abundance, defense conversion, and useful-work mining could fund a livable basic income. TL;DR: As impossible as decreasing military spending and closing corporate loopholes may seem, and as sappy as this sounds, Hope, Unity and Love are still stronger. It is

## Product

- [10 CI/CD Release Failures in Custom LMs — Which Stage Catches Each](https://divinci.ai/blog/10-ci-cd-release-failures-in-custom-language-models/): Ten real LM-release failure modes, each mapped to the Divinci pipeline stage — Register, Gate, Roll, Observe — that catches it before users notice. Notes from the Release Cycle — Part II --- The first post in this series walked through the four-stage release pipeline we ship — Register → Gate → Roll
- [Automated LLM CI/CD Pipelines With Instant Rollback](https://divinci.ai/blog/automated-llm-ci-cd-pipelines-with-instant-rollback/): The operational layer under the four-stage pipeline: which decisions fire automatically, what an actual rollback drill looks like, and the MTTR number. Notes from the Release Cycle — Part V --- The most-quoted page that didn't go out last quarter was the one our observer fired on its own at 2:14 AM.
- [Automated Regression Testing for Custom LLMs in 2026](https://divinci.ai/blog/automated-regression-testing-for-custom-llms-in-2026/): How to build a regression suite that catches drift in the eval — not just the model. Slice-aware gates, calibrated judges, production-trace replay. Notes from the Release Cycle — Part 7 Friday at 4:47 PM you shipped a one-character prompt tweak. The aggregate eval score moved from 0.873 to 0.871 — w
- [CI Testing for Custom Language Models in 2026](https://divinci.ai/blog/ci-testing-for-custom-language-models-in-2026/): Contract tests, smoke budget, cost-aware fleet sizing, and shadow CI. How to keep a 12-minute eval suite tractable on every PR without slowing the team. Notes from the Release Cycle — Part 8 (final) You ship the regression suite from post 7. It works. The slice-aware gates catch real bugs. The calib
- [Calibrating the Judge: The Grader get Graded](https://divinci.ai/blog/calibrating-the-ai-judge/): ScoredQA Calibration: a domain expert rates 50 answers, we compute Spearman ρ vs each LLM judge, and pick the judge that actually agrees. --- The whole product is two equations. For each candidate LLM judge $j$, with $n$ paired (human, judge) ratings on the same answers, compute Spearman's rank corr
- [How to Build an LLM CI/CD Pipeline With Divinci AI](https://divinci.ai/blog/how-to-build-an-llm-ci-cd-pipeline-with-divinci-ai/): A four-stage LLM release pipeline: slice-aware Spearman gates, canary on output quality, 12-second atomic rollback, compliance receipts per decision. Notes from the Release Cycle — Part I --- The first time we tried to ship an LLM through a normal CI/CD pipeline, the build went green, the deploy suc
- [How to Diagnose Custom LLM QA Failures in 7 Steps](https://divinci.ai/blog/how-to-diagnose-custom-llm-qa-failures-in-7-steps/): Most 'QA failures' aren't model failures — they're eval gaps, judge mis-calibration, or training-serving skew. A 7-step diagnostic that proves it. Notes from the Release Cycle — Part VI --- A scored-QA suite started flagging a customer's medical-Q&A model. The headline number — aggregate quality acr
- [The 12 QA + Release Capabilities Every Custom-LLM Platform Ships](https://divinci.ai/blog/12-qa-and-release-management-capabilities-for-llms/): Capability checklist for LLM release platforms: slice-aware gates, calibrated judges, atomic rollback, hash receipts — what ships, what's missing. Notes from the Release Cycle — Part III --- A year ago, before we started building our own release pipeline, we sat down and listed every QA-and-release

## Research

- [Deleting Paris from a Language Model](https://divinci.ai/blog/deleting-paris-from-a-language-model/): A vIndex patch deletes 'Paris is the capital of France' from a frozen LLM. The receipt is a 100-byte JSON file with a SHA-256 checksum. The Interpretability Diaries — Part II --- Before the edit: After a single rank-1 DELETE patch: That's Gate 3. A single weight-space edit, targeting one feature in
- [Inside the RAG Arena: When the Judges Don't Agree](https://divinci.ai/blog/inside-the-rag-arena-scored-qa-routing/): A 200-item RAG arena tied at the mean, but two LLM judges only agreed at Spearman ρ=0.55. They aren't measuring the same thing. The whole experiment is two equations and one footnote. Per-variant overall score , mean over all $S$ scorer rubrics applied to all $N$ scored items: $$ \overline{x} v \; \
- [Prominence Is Not Relevance: Teaching a Storefront Assistant What to Recommend](https://divinci.ai/blog/prominence-is-not-relevance/): An e-commerce assistant linked products correctly but recommended the wrong ones. The fix was discovering that decoration, ranking, and recommendation are three different problems — and that prominence is only ever a tiebreaker on relevance. The whole feature is one score and one ordering rule. The
- [Speculative Decoding for Free: Pairing DFlash with our DFO-Tuned Gemma 4 31B](https://divinci.ai/blog/gemma4-dflash-speculative-decoding-for-free/): z-lab's DFlash drafter on our QLoRA fine-tune captured 92% of the published speedup with no retraining. ~15x faster, ~4x cheaper in prod. --- TL;DR We have a fine-tuned 31B-parameter Gemma 4 served on Modal H100 — Direct Free-text Optimization (DFO), our internal SFT/DPO mix on the AskTheDoctor medi
- [The Architecture Every Language Model Converges To](https://divinci.ai/blog/architecture-every-llm-converges-to/): Every modern LLM converges on the same four-stage circuit. We map the architecture and what its absence in 1-bit models means. The Interpretability Diaries — Part I Updated April 23, 2026 — Kimi-K2 (Moonshot AI, 1T-param MoE) just dropped. It's the fourth independent organization to build a frontier
- [The Two Models That Never Met. Both Measured at the Same Depth.](https://divinci.ai/blog/two-models-never-met/): Two natively-trained 1-bit LLMs converge on the same activation anomaly without ever sharing weights. A note on convergence under pressure. The Interpretability Diaries — Part IV --- Two teams. Different organizations. Different training data. Different architectures. Different hyperparameters, diff
- [WWW-RAG: Making the Open Web Chattable, From One MacBook to Cloudflare](https://divinci.ai/blog/www-rag-making-the-open-web-chattable/): A Rust daemon on one MacBook published a dozen sites a day. The same pipeline now runs on Cloudflare, unattended, and publishes about a thousand. --- What WWW-RAG is Point your browser at the WWW-RAG directory and you'll find a shelf of real websites — the Internet Encyclopedia of Philosophy, NASA,
- [We Made Our RAG Pipeline Parse PDFs 20–50× Faster](https://divinci.ai/blog/liteparse-vs-openparse-faster-pdf-parsing/): We swapped OpenParse for LiteParse in our RAG ingestion pipeline. The headline is 20–50× faster parsing. The useful part is the four ways we got it wrong first. --- TL;DR We replaced the PDF parser in our RAG ingestion pipeline — from OpenParse (a remote Python service that converts PDFs to markdown
- [We Tested Headroom Against Our EXIT RAG Compressor](https://divinci.ai/blog/headroom-rag-context-compression/): Open-source Headroom compresses RAG context with an ONNX model. We wired it into our pipeline and raced it against a 50-line in-process extractor. Neither won outright. --- TL;DR We have a pluggable RAG context-compression slot — retrieved chunks pass through a strategy before they reach the model.
- [When the Circuit Dissolves](https://divinci.ai/blog/when-the-circuit-dissolves/): Two natively-trained 1-bit LLMs lose the four-stage circuit that organizes fp16 transformers. Behavior survives; structure dissolves. The Interpretability Diaries — Part III --- The 1-bit models still answered correctly. Their internals had no structure. Then I ran the second one. Same result. --- A

## Technical Guides

- [Optimizing Vector Embeddings for Better Search Results](https://divinci.ai/blog/optimizing-vector-embeddings/): Practical techniques for tuning vector embeddings: dimensionality, quantization, and retrieval quality trade-offs we use in production. Vector embeddings have become the backbone of modern AI-powered search and retrieval systems. From RAG applications to recommendation engines, the quality of vector

## Vision

- [Light Logic: The Dawn of Photonic Ternary Computing for Universal Basic Compute](https://divinci.ai/blog/light-logic-ternary-computing/): Photonic gates, ternary computing, quantum RNG, gravity batteries — a sustainable compute substrate for UBI. May 2026 — three of four pillars commercial. About this post. This is a vision piece: photonic logic gates, ternary computing, quantum RNG, and gravity batteries are real research areas, but