About
Founder building Super Creativity series at Enlighten Animation Labs. I learn through…
Activity
4K followers
-
Bell C. shared thisCan Claude Code beat Ahrefs Agent A? Only if the moat is workflow. Not if the moat is live data. This is the part I think SEO teams, agencies, and founders should care about. Agent A is not interesting because it has a chat UI. A custom Claude Code harness can already rebuild a lot of the workflow if you feed it: - Ahrefs exports - Google Search Console - page URLs - traffic decay - old briefs - brand rules It can inspect schemas, cluster intent, rank refresh opportunities, and save the process to run again next week. The harder thing to copy is Ahrefs' live graph: backlinks, keyword history, SERP movement, content explorer, and brand visibility signals. So my test for vertical AI agents is simple: If the value is reasoning over your files, build the harness. If the value is live proprietary data, buy the vertical agent. The agent layer is getting cheap. The data layer is still expensive. Would you trust a Claude Code SEO harness over Agent A if all your data were connected? #AIAgents #SEO #ClaudeCode #ContentMarketing
-
Bell C. shared thisMajor update to Superdirector: it now starts with a brand profile URL, not a blank prompt box. That sounds small. It is not. Most AI tools for content teams start too late in the workflow. They help after you've already decided what to make. But for most social media managers, that isn't the hard part. The hard part is: - figuring out what is actually working in your niche right now - knowing whether it fits your brand - turning that into something your team can actually shoot this week So we rebuilt the core flow around this: 1. Paste a brand or creator profile URL 2. Get an automatically generated niche-specific viral feed 3. Pick one scored idea 4. Turn it into a script and shot plan The goal is not "generate more content." The goal is to make content strategy feel operational. If you run social for brands, creators, or agencies, I'd love to show you what we're building. Your journey starts here: https://lnkd.in/gcUw8zih #socialmediamarketing #contentstrategy #brandstrategy #contentplanning #socialmediagrowth #socialmediamanager
-
Bell C. reposted thisBell C. reposted thisWe are launching our product: the multi-member team content for B2B companies! Most companies try LinkedIn GTM with one founder posting alone. It works, but it's slow. One voice. One network. Limited reach. What actually moves the needle is coordinated team effort. Multiple executives posting aligned content, amplifying each other, creating a ripple effect across the org. That's what we built at Imagine AI (YC F25). How it works: We go through your transcripts, interviews, posts, and conversations. From that, we build a 100+ page Persona Report for each person on your team. It captures how you actually think and talk. That's why the content sounds like you wrote it. Not a polished version of you. You. One calendar for your whole team. Your CEO posts the vision, your VP of Sales posts the proof, your Head of Marketing posts the framework. Same theme, different angles, coordinated cadence. Every client gets a dedicated Content Engineer who owns your strategy and has an AI platform behind them. Results across 40+ B2B companies: -> $31M in revenue generated -> 431 qualified inbounds in 1 month for one client -> $53K inbound revenue in 1.5 weeks for another -> 12M total impressions -> 10,400 hours saved We built this because our customers kept asking for it. They wanted their whole team posting, but couldn't justify the time investment for each person. Now they can. We put together a breakdown of the exact content framework we use with our clients to hit Content Market Fit. Comment "CMF" and I'll send it to you. imagineai.me
-
Bell C. posted thisAI can generate stunning videos now. But when I talk to creators, they're not asking for more realism. They're asking: "Why did my last Reel flop when the one before went viral?" "What's the right pacing for a 30-second video?" "How do I structure a hook that actually stops the scroll?" The generation tools keep getting better. And yet the creative questions remain the same. What to make. How to structure it. Why one video connects and another doesn't. The patterns are surprisingly consistent: → The hook earns the next 3 seconds → The pacing matches the emotional arc → The structure makes even short videos feel complete These aren't technical problems. They're craft. And they've always been what separates good content from forgettable content—long before AI entered the picture. The tools for making videos are incredible now. I'm curious what tools will emerge for understanding them. What's the hardest part of your creative process right now? #CreatorEconomy #ContentStrategy #UGCMarketing #Storytelling
-
Bell C. shared thisTwo months ago, I was watching my social media marketing interns spend 3+ hours on each video breakdown. Not creating. Not shooting. Just PLANNING. They'd watch a viral video 20 times, scribble notes, try to reverse-engineer the viral shots, guess the techniques, and still end up with bad hooks. It was painful to watch. So I built a tool. Drop a reference link. Get a complete shoot plan — shot lists, scripts, gear recommendations, lighting setups, even storyboards, all in your budget and time frame. Superdirector is your personal AI director that shortcuts everything between exploring ideas and picking up camera in minutes. What used to take my interns 3+ hours now took 3 minutes. I thought it would stay internal. Just something to save my team time. Then my filmmaker friends started asking for access. The same week, three different people asked the same question. That's when I knew this wasn't just our problem. The gap between "I love this video" and "I can actually recreate this" is massive. Most creators never bridge it. They get stuck in the planning phase forever. Superdirector bridges that gap. What started as an internal automation tool is now helping social media managers, filmmakers, UGC creators, and film students turn inspiration into execution. I built this because watching my team struggle was personally painful. Turns out, that pain was more universal than I realized. Have you ever built something for yourself that ended up helping way more people than you expected?
-
Bell C. reposted thisBell C. reposted thisImagine AI (YC F25) builds B2B content pipelines by cloning you. They create high-fidelity personas from your expertise, then handle posts, comments, and category-defining blogs that drive pipeline. What Imagine AI is actually doing is content engineering to get you to content market fit. They optimize based on engagement data until your content consistently drives qualified inbounds. The results? 10M driven across clients. One founder landed a $5M deal from a single post. Another got 200 qualified inbounds in a month. Founded by Sky Yang (raised $150M while at UC San Diego) and Neo Lee (Berkeley AI researcher who dropped out to build this). With Imagine AI, they're building the content OS for every LinkedIn-native B2B company. Book a demo at imagineai.me or reach out at sky@imagineai.me. Congrats on the launch, Sky and Neo! 🚀 https://lnkd.in/gh_zWBBU
-
Bell C. shared thisIt's really a pleasure to design a pitch deck using the product that I am pitching and I can feel the breathing of multimodal AI behind the canvas.
-
Bell C. shared thisCongrats Ziqi Huang on this amazing paper! This is a big step for world modeling leveraging multimodal reasoning in visual generation.Bell C. shared thisVideo generation has come a long way, but crafting complex scenes with a truly coherent chain of events remains a significant challenge. A new paper, "VChain: Chain-of-Visual-Thought for Reasoning in Video Generation" by researchers from Eyeline Labs and Nanyang Technological University, introduces an exciting solution. VChain is an inference-time framework that injects powerful visual reasoning from large multimodal models, like GPT-4o, directly into video generation. It works by having the multimodal model generate a sparse set of critical "visual thoughts" – keyframes that capture the essential causal steps of a scenario. These keyframes then guide a lightweight, sparse tuning of a pre-trained video generator. The result? Videos that not only look smooth but also logically follow a sequence of consequences, enhancing dynamic fidelity and narrative coherence without extensive retraining or dense supervision. This is a game-changer for creating more realistic and intelligent video content. Want to explore the details of VChain, its innovative framework, and see compelling examples? Check out the paper, code, and project page below: Paper: https://lnkd.in/eUBtxwPs Project Page: https://lnkd.in/eXNknZjc Code: https://lnkd.in/en_Nh22W Video Demo: https://lnkd.in/e6v65iaW
Bell C. commented on a post
4h
My sharper framing after thinking about this:
For marketers, the practical question is not "which agent is smarter?"
It is: what data can the agent see that my own harness cannot?
If you run SEO/content, where would you draw the build-vs-buy line?
-
Bell C. liked thisBell C. liked thisJust got off a call with Clavicular about going viral, here’s his insights: 1. Stop trying to be everywhere Double down on the one channel where your audience already pays attention. 2. Clear messaging beats clever messaging. If people don’t instantly get it, they won’t care. Even Steve Jobs’ best speech only talked about one thing. 3. Get people’s attention first, and explain the details later. This is why this post is fictional and generated by ChatGPT to promote Ditto’s blind date yacht party at Boston on May 3rd Sign up now, link in comments
-
Bell C. liked thisBell C. liked thisAt 17, I was done being a shy kid. By 40, I built an AI compute company to $120M ARR from my basement. Hi LinkedIn, I'm Zhen! Here was everything that happened between 17 and 30: 1. High school - Played StarCraft semi-professionally but could barely hold a conversation with anyone in real life 2. Undergrad - Had a pretty rough time initially (50% of my freshman class dropped out and got my first C in a class) - But enjoyed the period of self-experimentation: took drama courses, did improv, and dyed my hair red 3. Grad school + teaching - Loved physical chemistry in undergrad so much that I continued pursuing it with a PhD at Temple - But teaching felt too slow, and I ended up quitting my professorship after two years - Career wise, it was back to the drawing board. I taught myself how to code and started interviewing for software engineering jobs 4. Comcast - Lucked out when Pardeep took a chance on me and hired me to join his team - We worked together for six years building distributed systems, a 100+ person team, and an incredible friendship - In 2022, we decided to bite the bullet and start a company 5. Runpod: early days - First MVP was trash: just servers running in my and Pardeep's New Jersey basements powered by Comcast home Internet - The company logo was so generic that scammers would use it to phish our customers - Slowly but surely, over the course of 9 months, we crawled our way to $1M in revenue 6. Runpod: now - We've grown from 2 to 90 people - We just blew past $120M ARR, have 500K+ users worldwide, and power infrastructure for frontier AI labs and Fortune 500 companies ...and we're not stopping there! Our goal: to be the only compute platform that growing AI companies need.
-
Bell C. liked thisBell C. liked thisMy PhD was supposed to be about oil and gas emissions. And it is — but it accidentally also became a crash course in why LLM inference is broken at scale. I was running thousands of LLM calls to extract operational parameters from technical documents. The models worked. But the costs kept climbing in ways that didn't make sense. Every request was reprocessing the same system prompt, the same context, the same examples — from scratch, every time. That frustration is what pulled me toward SGLang and RadixArk. The core idea behind SGLang is pretty simple once you see it: if you've already computed something, don't compute it again. RadixAttention does this by using a radix tree to identify shared prefixes across requests and reuse the cached KV tensors instead of recomputing them. At scale — across thousands of concurrent users — the savings are real. We extended the same logic to image generation. Diffusion models have the same redundancy problem as LLMs, and the same caching + parallelism strategies apply. I'm excited to share that our free course is live on DeepLearning.AI and this was built in partnership with LMSYS and RadixArk A huge thank you to Andrew Ng and the DeepLearning.AI team. This collaboration meant a lot, and I hope it helps everyone can actually understand what's happening under the hood when they deploy models in production. If you're running LLMs at any scale, this stuff matters more than most people realize. 👇 https://lnkd.in/gfdaghZd
-
Bell C. liked thisBell C. liked thisLast week I gave a talk to 100+ founders at Workato and Bond AI about how we think about B2B growth at Imagine AI. Huge thanks to Yelaine Wang for organizing and moderating. She put together something really special. Here's the core idea I shared. The B2B landscape has completely shifted in the last two years. AI made content volume explode. Everyone can write now. Everyone can send outbound now. Messages that used to be static and templated are now hyper-personalized and dynamic. But here's the problem. The volume went up across the board. Your inbox is flooded. Your LinkedIn DMs are flooded. Your Gmail is flooded. Be honest. Do you actually read your LinkedIn messages anymore? Do you read cold emails? Most people don't. Now imagine this instead. You've engaged with someone's content four or five times. You've seen their takes. You've commented on their posts. You've liked their stuff. Then they send you a message. You respond. Of course you do. You already know them. B2B growth has always been about trust. It's always been about value. It's always been about familiarity. That hasn't changed. What changed is how you build it. The companies winning right now aren't the ones blasting more messages. They're the ones whose executives and GTM leaders are posting consistent, coordinated content on LinkedIn. Building trust before the first sales call ever happens. That's what we're building at Imagine AI. We want to be the company that systematizes B2B growth the correct way. Not more noise. More trust. At scale. Huge thanks to everyone who made this happen: Francis Li, Sahar Mor, Emily Fang, Elly L. [Created by Imagine AI]
-
Bell C. liked thisBell C. liked thisThe New York Times just wrote about me leaving a Chinese New Year party early to check on my AI agents. That sentence alone tells you how much the world has changed. A year ago, software was the expensive part. You hired engineers, you built slowly, you waited. Now software is cheap. I run 5 AI agents simultaneously using Anthropic's tools. I upgraded from $20/month to $200/month because the leverage is that insane. Any founder can spin up agents that code, research, draft, and execute. So what's actually expensive now? Knowing what to build. Knowing who to build it for. Knowing why it matters. Direction. Creativity. The human layer. This is exactly why the content landscape is broken right now. AI can mass-produce blog posts, newsletters, and social content at a rate nobody has ever seen. And most of it is slop. Everybody can make a blog now. Everybody can generate a newsletter. So how do you stand out? You don't stand out with volume. You stand out with voice. That's why LinkedIn exploded over the last two years. After COVID, after GPT got good, people realized the only content that cuts through is personal. It's the founder behind the company. The human behind the brand. B2B buyers want to know who you are before they ever get on a call. Trust is the new moat. And trust comes from a real person with a real perspective. This is what we're building at Imagine AI (YC F25). Not another content tool. A content engine that captures the one thing AI can't fake on its own: you. Your voice. Your stories. Your point of view. At scale. The era of generic content is over. The companies that win from here are the ones whose prospects already trust them before the first meeting. Link to the full NYT article in comments. [Created by Imagine AI]
Experience & Education
-
Enlighten Animation Labs
View Bell’s full experience
See their title, tenure and more.
or
Publications
-
Generate Desired Images from Trained Generative Adversarial Networks
IEEE International Joint Conference on Neural Networks
The emerging of Generative Adversarial Networks (GANs) gives rise to a significant improvement in image generation. However, a controllable way of synthesizing images with specific characteristics still is a challenging issue. Many existing methods are not efficient enough that require additional information and pre-designed attributes, and are with much more human intervention. In this paper, we propose GAGAN, an extension method to the Generative Adversarial Network, which is the first work…
The emerging of Generative Adversarial Networks (GANs) gives rise to a significant improvement in image generation. However, a controllable way of synthesizing images with specific characteristics still is a challenging issue. Many existing methods are not efficient enough that require additional information and pre-designed attributes, and are with much more human intervention. In this paper, we propose GAGAN, an extension method to the Generative Adversarial Network, which is the first work to generate specific images from a trained GAN model. To control the characteristics of images, a DNA pool of the trained GAN model is introduced and evolved by a genetic algorithm (GA). Then, with the DNA pool, GAGAN can generate the corresponding latent vector (DNA) of target images. Furthermore, GAGAN can synthesize images containing a single specific characteristic or multiple specific attributes (including AND and OR relation). Moreover, several fitness evaluation strategies are also proposed to make GAGAN flexible to control the target characteristics. Experiments on CelebA and MNIST are conducted, and results show that the proposed method is feasible and effective in specific image generation problem.
Other authors -
Projects
-
Soft Actor-Critic learns High-Frequency Crypto Trading
-
See projectCollected and created backtest environment from datasets of 5 crypto trading pair spread. Trained small SAC model to frequently trade between pair that maximize return rewards. Experimented design of logarithmic reward function.
-
LSTM for Code Completion
-
Trained next-word prediction LSTM for C/C++ code completion. Data scrapped from GitHub, containing ~50 large code library of C/C++. Created a simple IDE with QT UI.
-
Improved 3D U-Net and CycleGAN for Character Rendering
-
See projectCreated efficient async GPU data loader for video training. Experimented new 3D U-Net architecture in CycleGAN for skeleton rendering and style transfer task. Designed synthetic data generation pipeline for rigged 2D characters.
-
IEC Device: Trial of Nuclear Fusion Reaction
-
See projectModified the design of original FUSOR device and constructed demo nuclear fusion device. Tested with Deuterium gas with 1e-5 bar pressure glow discharge at 10-15 kv.
Honors & Awards
-
Second Place in Ohio Hack AI Competition
OHI/O and Artificial Intelligence Club
Project: Deep-learning-based video depth prediction
-
Amaze Design Price in VEX Robotics World Championship
REC Foundation
-
First Price in Provincial Innovation Competition
Sichuan High-school Technology and Innovation Competition
Project: Machine learning EEG oral recognition for the mute
Other similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content