Thomas Chong

AI Research Engineer @Beever_AI | @GoogleDevExpert in Cloud AI | HKUST | 🇨🇦

[Beever AI @Beever_AI]

Two weeks ago @karpathy posted about "LLM Knowledge Bases" — the idea that LLMs should maintain structured, evolving knowledge from your documents. 1.7M+ views. He said:

"I think there is room here for an incredible new product instead of a hacky collection of scripts."

We've been building exactly that. Today we're open-sourcing Beever Atlas. GitHub: https://t.co/61A5VbpXln

The difference: Karpathy's approach starts with manual file uploads. Beever Atlas starts with your team's chat. Slack, Discord, Teams, Telegram — the messy, unstructured conversations where 90% of organizational knowledge actually lives and dies.

Here's what it does:
- Connect your chat platform (self-service, takes 2 minutes)
- Ingestion pipeline extracts entities, facts, and relationships automatically
- Builds a Neo4j knowledge graph — not just text cross-references, actual typed relationships between people, projects, technologies, decisions
- Generates a living Wiki — DeepWiki-style, with topic hierarchies, concept maps, glossaries. Updates every sync.
- Ships as an MCP server — Claude, Cursor, any AI assistant can query your team's collective knowledge directly

From our internal deployment (4 Slack channels):
- 854 structured memories
- 1,899 entities
- 5,271 relationships
- 222 wiki entries auto-generated

What Karpathy built is single-user, requires Obsidian + CLI, text-only. Beever Atlas is multi-user, zero-install web UI, knowledge graph, MCP-native.

We built this at Beever AI, a Toronto-based research lab under Votee AI, because we needed it ourselves — our engineering team's context was scattered across Slack threads nobody reads. Now our agents can actually reason over what the team knows.

100% on-premise. Docker stack. Bring your own LLM via LiteLLM (Ollama, Gemma 4, whatever you run locally). Zero data leakage.

Turn your team's chat into a living wiki.

⭐ https://t.co/61A5VbpXln
💬 https://t.co/XzLTuPyWXy
🌐 https://t.co/pQf9RCfjXc

Shipped by the whole team:
Engineering — @jhkchan @cch_thomas @KaiYamYang1 @dantelok1111
Design — Adrian Leung
Comms & Media — @nghoihin

Beever AI is a Toronto-based research lab under @Votee_AI.

[Tejas Gawande @tejgw]

Cursor for Slides is finally here
Watch the first 47 seconds. Then try going back to your old deck tool

Reply "Chronicle" + RT to get two months of Pro for free. Make sure you follow so I can DM you asap.

[Thomas Chong @cch_thomas]

@tejgw Chronicle

[Junyang Lin @JustinLin610]

me stepping down. bye my beloved qwen.

[Thomas Chong @cch_thomas]

@OfficialLoganK Would like to test it 💯

[Max Ku @vinesmsuic]

We tried to stress-test current image generation models for every possible real-world task.
What we found reveals a lot about how today’s models see, edit, and fail.
🧵 (1/N)

[Kevin Lin @KevinQHLin]

Thanks @_akhaliq for sharing our work!!

🤔Today’s video generation models (e.g., Veo3, SoRA) are great at realism, but they still struggle to convey structured knowledge and logical teaching.

🌟Code2Video🌟takes a different path: starting from Python Manim code, it renders project-level programs into educational videos—bridging coding, visualization, and knowledge!

📷 Code: https://t.co/gb0JeOTHv8
🏠 Website: https://t.co/8sDvaoEHqm
📄 arXiv: https://t.co/uFW2l18z8e

We want to share our gratitude to @3blue1brown and @manim_community !!!
Thanks to the great team @Anno_YanzheChen and @MikeShou1 !

#VIDEO #education #Sora2

[Thomas Chong @cch_thomas]

I will be attending 𝗔𝗖𝗟 𝟮𝟬𝟮𝟱 𝗶𝗻-𝗽𝗲𝗿𝘀𝗼𝗻 (27/7 - 1/8) to give an 𝗢𝗥𝗔𝗟 𝗽𝗿𝗲𝘀𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 on our paper, TheoremExplainAgent (https://t.co/IKywLiaWnz)

I warmly invite you to join my presentation to see our work in action. 👇

▶️ 𝗦𝗲𝘀𝘀𝗶𝗼𝗻 𝗡𝗮𝗺𝗲: 11 - IP-Orals, Multimodality and Language Grounding 2
▶️ 𝗧𝗶𝗺𝗲: Wednesday, July 30, 09:00 - 10:30.
▶️ 𝗟𝗼𝗰𝗮𝘁𝗶𝗼𝗻: Hall M2, Austria Center Vienna

On the other hand, I will be at the venue for the duration of the conference and would love to connect! 🤝 My interests are AI Agents 🤖, Multimodal LLMs 🗣️🎨, and Low-resource Languages 🌍. If you're working in these areas or just want to discuss, please reach out! 👋

See you in Vienna!

#ACL2025NLP #AI #NLP #Research #LLMs

[Thomas Chong @cch_thomas]

Just watched #GoogleIO & immediately applied for the Gemini Diffusion waitlist. Amazed at how quickly I got approved! 🤩

Tested it by asking for UI for an internal tool to fetch/summarize research papers (arXiv, Hugging Face, Slack) & plan project integration. The generation speed is mind-blowing (1044 tokens/s)! 🤯

Code quality might not be as good as Gemini 2.5 Pro/Flash, but it's fantastic to see how fast for rapid prototyping.

Huge potential for Diffusion LLMs in iterative tasks like coding & writing. Excited to see it evolve!

#GoogleIO #GeminiDiffusion #AI

[Thomas Chong @cch_thomas]

🎉 HUGE NEWS! 🎉
My first paper, "TheoremExplainAgent" has been ACCEPTED to ACL 2025 main (@aclmeeting)! 🚀

So proud of what we have accomplished. We explored using AI agents to generate long-form theorem explanation videos!

Can't wait to share our work in Vienna! 🇦🇹

@vinesmsuic Johnathan Leung @KrishRShah Alvin Yu @WenhuChen

@HelloVotee

#ACL2025 #TheoremExplainAgent #NLP

👇 Check out our original work!

[Max Ku @vinesmsuic]

🚀 We just released the code for #TheoremExplainAgent! 🧮🎬

Agentic LLM approach generates long-form (>5 min) theorem explanation videos using Manim. While highly successful, layout issues remain. We also introduce TheoremExplainBench for systematic evaluation.

👇 Details (1/n)

[Hung Chia-Yu @hungchiayu123]

🚨 Big News 🚨
Our paper "Inference-Time Alignment with Reward Models" has been selected for an oral presentation at #NAACL2025! 🎉
🗓 May 1 | ⏰ 14:30–14:45
📍 Ballroom C (Session H)

If you're into inference-time alignment and reward-guided decoding, don't miss this! 🔥
Sadly, I won’t be there — but my amazing advisor @soujanyaporia will be presenting on our behalf! 🙌

[Votee App @HelloVotee]

Our AI research team (@cch_thomas) published TheoremExplainAgent paper🚀! This paper explores an agentic approach to generate long-form videos w/ Manim animations to explain theorems.

Website: https://t.co/DtWeTzq05i
GitHub: https://t.co/5CJlrR4hsl
arXiv: https://t.co/ZfyABLTkC0

[Thomas Chong @cch_thomas]

Big thanks to #TheoremExplainAgent team: Max Ku @vinesmsuic, Jonathan Leung, Krish Shah @KrishRShah, Alvin Yu, and @WenhuChen.

And most importantly, @HelloVotee for supporting this project!

Special thanks to @_akhaliq and other influencers for promoting the paper.

[7/7]

[Thomas Chong @cch_thomas]

I'm excited to share my first publication as a co-first author : #TheoremExplainAgent! 🎉

We use agentic LLM approach to create >5 min long theorem videos w/ Manim. TheoremExplainBench is also released for systematic evaluation.

👇 [1/7]

[Thomas Chong @cch_thomas]

We have just released the code and feel free to know more details below:

📌 Project Page: https://t.co/oS7MxQNBjj
📑 Paper: https://t.co/SiBoCQPQT2
⌨️ Code: https://t.co/cpOmw403dS

[6/7]