Graphon AI

At Graphon, we believe intelligence lives in the connections, not the nodes. So we are bringing the graph to the table. We are hosting our first Engineering Leaders Dinner sponsored by J.P. Morgan with confirmed attendees from OpenAI, Anthropic, Harvey, Carta, Google, Hippocratic, Lucid Motors, Prime Intellect, Mercor and more. We have a few seats left and if you are passionate about topics like context intelligence, memory, knowledge graphs, multimodal data and LLMs would love to have you join 👇 luma.com/k7nnf5g6 #MultimodalAI #EngineeringLeaders #Agents #ContextIntelligence #Memory #Graphon

Graphon is hiring in San Francisco. We recently emerged from stealth with $8.3M in seed funding and we're growing fast. We’re looking for builders, researchers and storytellers excited to shape what comes next for enterprise AI. Open roles: Full Stack Engineer, ML Infrastructure Engineer, Research Engineer, Founding Marketer and Interns. Apply + share: job-boards.greenhouse.io/graphonai

No matter the enterprise AI use case, the same two flaws exist underneath every one of them: logical blindness and memory decay. @deepakmishraVC explains why bigger models won't fix it.

AI isn’t blind. It just can't hold enough of your data at once to reason over it. And this is a foundational limitation. @ArbaazK87777726 on why enterprises need a pre-model intelligence layer to help their models go further.

Graphon is out of stealth today. $8.3M in seed funding. We’re building the pre-model intelligence layer — the layer that sits before your AI model and maps how data connects. Every model has a ceiling on what it can process. Your data shouldn’t. More in @WSJ: wsj.com/tech/ai/grapho… Learn more: graphon.ai

This is a really clear way to think about context windows. The “lost in the middle” problem, how RAG tries to compensate, and why people end up resetting instead of building on prior context all connect here. At some point, it’s not just about how much information you include, but whether that information holds together in a way the model can actually use.

This article from @techreview gets at something we’ve been thinking about while building Graphon. Companies have Step 1 (build the tech) and talk about Step 3 (transform the business). But Step 2 — making it work in real workflows with real data — is where things slow down. In one study, AI agents were tested on 480 everyday tasks and failed most of them. That’s exactly the problem Graphon is working to solve. technologyreview.com/2026/04/27/113…

@Stanford's 2026 AI Index is out ➡️ Multimodal publications are up 2.7x in two years, and models are clearing benchmarks almost as fast as they're written. But it also reveals that these same models can't read an analog clock. This is an architecture problem. Multimodal is moving fast, but the infrastructure behind it is still catching up. That’s exactly the gap we’re closing with Graphon. More to share soon! spectrum.ieee.org/state-of-ai-in…

Enterprise AI buying has gotten harder, and the data shows it. Only 11% say deployments are meeting core goals. Models work. Systems don't. Retrieval finds data but can't explain relationships, so reasoning stays shallow and ROI stays out of reach. The bottleneck isn't capability. It's what happens before the model runs. cc: @WSJ wsj.com/articles/selli…

AI spend is concentrating. A few vendors win, and budgets shift fast. But the foundation hasn’t changed. More model spend ≠ better outcomes. The gap isn’t capability. Its structure. That’s what the next layer solves. techradar.com/pro/security/a… cc: @TechRadar

This is exactly the product we have been building at @graphonai. Instead of compiling raw data into .md and re-reading it into a context window, you build a permanent relational map the AI reasons over directly. Scales far beyond the ~400K words, and across videos, audio, images too. We call it a persistent relational memory. Happy to give you API access to try it on your wiki.

LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

Documents were just the beginning. Most enterprise data lives in video, audio, and mixed formats…and current AI can’t reason across it. @ArbaazK87777726 on why multimodality is the next real unlock. #EnterpriseAI #MultimodalAI

AI is moving fast. The data layer isn’t. Vector DBs find similarity. SQL organizes data. But neither helps AI reason across relationships, context, and modalities. That missing layer is the unlock for agentic systems. That’s where Graphon comes in. dell.com/en-us/blog/ai-…

AI brittleness shows up in the smallest places. Ask a simple follow-up question, and the system breaks. @ArbaazK87777726 explains why this isn’t a tuning issue; it’s an architectural one. #EnterpriseAI #MultimodalAI

Enterprise AI isn’t blocked by ambition. It’s blocked by a gap between use cases, value, and multimodal data. @deepakmishraVC on the real-world challenges enterprises are encountering — and why architecture matters. #EnterpriseAI #MultimodalAI

AI brittleness shows up in the smallest places. Ask a simple follow-up question, and the system breaks. @ArbaazK87777726 explains why this isn’t a tuning issue; it’s an architectural one. #EnterpriseAI #MultimodalAI

Enterprise AI has a missing layer. Between ingestion and copilots. Between storage and reasoning. Petabytes in. Models on top. No durable long-context infrastructure in between. That gap is now the strategic layer. Graphon fills it.

There’s a hidden tax in AI infrastructure: Teams pay 10–50x more to dodge multimodal context limits. Stacking bigger models + more RAG ≠ fixing the constraint. When cost and complexity spike, it’s usually architectural. Graphon is the ultra-long-context intelligence layer built for multimodal enterprise data. If you’re seeing this, find out more: graphon.ai

This isn’t a scaling issue. It’s a mismatch between current architectures and enterprise reality. Enterprise knowledge doesn’t live in prompts — it lives in dense, multimodal systems.

Cross-modal reasoning breaks at enterprise scale.

Enterprise AI isn’t failing — it’s hitting an architectural wall. Not a talent problem. Not a data problem. Legacy AI stacks were built for text, not multimodal enterprise reality. Most AI stacks were built for text. Now they’re being pushed to reason over massive volumes of video, sensors, mixed media, documents, and edge cases. 🧵And then we’re surprised when: