Cofounder @RumiLabs_io | Prev adventures: COO @OmniFDN @harvardHBS @BCG @Amazon @Spotify @IBMwatsonx
Manhattan, NY
Joined August 2017
- Tweets90
- Following386
- Followers602
- Likes502
Some news that's been weighing on my heart: I've stepped down as Co-founder and President of Rumi Labs. A thread on what we built, and what's next. 🧵
To my co-founder Niko, with Rumi now fully in your hands: the vision is real. I believe Rumi is built to lead this revolution. Rooting for you!
Congrats @DAnconia_Crypto and @MoneroMahesh on closing Venture Fund II 🙌
Since co-leading @RumiLabs_io’s pre-seed, @EV3ventures has supported us far beyond capital: always showing up w/ real conviction & hands-on help.
Grateful to have you in our corner, and excited for all the founders you’ll back next
Today @RumiLabs_io is opening the gates for everyday computers to join the 1st-ever distributed Vision-Language Model network for Media.
We’re launching the Rumi Desktop App, running our proprietary SOTA edge VLM optimized for understanding complex multi-modal narratives.
If you want to be early on Vision Mode and help build the Narrative Intelligence Layer for AI, download the Rumi extension at rumi.io
This architecture allows Rumi to create the richest index of each media title. By pre-processing the video on their machines, Rumi’s nodes contribute to our vision of interactive media.
The index (exposed via APIs) allows AI agents to understand content and interact with it without the need to process video at each query. Instead of taking minutes to answer a single prompt, agents can interact with content in real time, with negligible retrieval cost.
With the Rumi Desktop App launch, we’re getting one step closer to the future where every video becomes an interactive, personalizable and shoppable experience.
Join the VLM network at rumi.io
Rumi's research paper is live! 📜
We're providing the 1st ever architecture allowing AI to transform any video into an interactive, personalizable, and shoppable experience.
And @RumiLabs_io is already beating SOTA VLM narrative understanding 👇
In preliminary evaluation on open-ended questions requiring multi-hop reasoning across 10-40 minute temporal spans, Rumi achieves 10% higher average score and 76% win rate against Gemini 2.5 Pro with raw video access.
Performance is strongest on temporal integration, multi-hop reasoning, and symbolic understanding tasks.
Check out our litepaper to learn more about our first-of-the-kind architecture and how we're able to compete with SOTA AI labs:
drive.google.com/file/u/1/d/…
Rumi envisions the future where every video becomes interactive, personalizable & shoppable
But existing tech cannot support our vision: AI lacks context, VLM querying takes minutes, and content we watch is behind IP wall
Solution: @RumiLabs_io Video Intelligence Architecture🧵
3. Shazam-like content identification:
To provide truly conversational experience for video interactivity (which we call “Friend talk”), we equip AI agents using Rumi’s Video Index with awareness of the current viewing moment
We developed audio fingerprinting technology that equips AI Agents with temporal grounding, allowing them to identify the precise timestamp location of the current playback position - think “Shazam for Media”
This mechanism makes interactions with AI agents feel like a naturally flowing conversation with a friend sitting on the couch next to a user and seeing what they’re seeing