Alexandria

Alexandria turns a person's videos, audio, and writing into someone you can talk to. Ask Steve Jobs a question, hear him answer in his own voice, grounded strictly in his own words, with sources.

Comment 3

Inspiration

Every subsequent generation of humanity is build upon the previous ones. The issue is, not all information is captured and then later realised.

Humanity has always built on the knowledge of the generation before it.

As funny as this sounds, we’ve literally gone from trees to wifi. The thing is, human knowledge accumulates but a lot of it is lost to time.

As technology advances, we continue to build on higher and higher levels of abstraction but what if we need to revisit a distant part of the chain?

Steve Jobs once said that Alexander the Great had Aristotle as his personal tutor which frustrated him to the moon because the closest he could ever get to learning from Aristotle was reading the texts he left behind.

But what if Aristotle could still answer questions? What if you could ask anyone a question. Imagine whose brain you could capture.

What it does

Alexandria ingests the data on someone (video, audio & text), provisions metadata and embeds the data for RAG indexing. Then, a user can ask someone like Steve Jobs a question through the predictive next-token inference of LLMs through informed and grounded knowledge of Steve.

The difference between a standard LLM vs Alexandria is the adherence to context and source grounding. It’s not “sound like Steve Jobs”, it’s “you are limited to the information of Steve from what has been provided to you and you must source it.”

How we built it

We spent the first two hours researching. We made a PLAN.md and then worked with AI to make a MVP.md which was a checklist of the things we needed to get done. We structured it based on getting to PoC as soon as possible so video avatars, which we didn’t end up touching, were last on the list just in case of time.

The codebase architecture was mapped to ensure: AI Readability = Human Readability. The reasons for this change are, (1) it’s the future, (2) human validation is still ESSENTIAL and (3) Majority Debugging & Test Write Ups could be delegated to AI.

At one point, latency wasn’t the best for the voice mode so we needed to create a LATENCY-PLAN.md file and have AI diagnose why.

We found some inspiration for UI pieces online which enabled us to rejig them into things we could use for the demo.

Challenges we ran into

Ambition made us have to redirect. Initially, we wanted to extend the project by adding a video component but we were running out of time and had to trash that.

The hardest thing we encountered was having to research and understand what was out there, set up Openrouter, Elevenlabs, Vercel, and THEN read the documentation before beginning to get to work.

Our goal was to build something that doesn’t take minutes for an answer like reasoning models. Finding the fastest budget friendly tools was manually intensive because even with AI assistance for searching it didn’t give the best tech choices that we ended up going with like Cerebras having a WAYY higher TPS than any other model provider for Gemma 4 31b.

Asynchronous sleep schedules did not work to our advantage. One of us was tired in the evening and one of us was tired in the morning, thank you coffee and sardines (yes really).

Accomplishments that we're proud of

We're extremely proud of the quality we were able to produce for the time constraint. Typically, there's a steep trade-off with the two but that's to us planning and coordinating, we really made a difference in what two students can do.

What we learned

Having only two people on a team for the idea we chose to materialise it, demo it, film and edit it is NOT RECOMMENDED. We realised an hour in that we had to map out the whole process to then divide and conquer. You don’t really think about it but chemistry amongst a team is really important; we were “reading” each other’s minds every now and then because we knew what we needed to do.

Trust your teammates. You might think you’re awesome at video editing yourself, for example, but you might end up shocked at how well they are at it.

What's next for Alexandria

Customer Experience: What AI Steve taught us during our testing is that customer experience is where you start, not where you finish. Engineering means nothing if no one wants your product. USyd, and Jane Street too, are very diverse in their reach regarding what industries they connect to. We think the best focus is on collaborating with USyd and Jane Street, if they’ll have us, to identify the best use cases through their large corpuses of market research and understanding.

Areas that strike out to us are:

  • Cloning Teachers: Teachers cannot offer 24/7 1-on-1 support but, with this technology, we can create a shadow tutor version of them with the entire course syllabus and material uploaded.
  • Cloning Industry Experts: A lot of expertise is lost when people retire, transition, etc. On top of that, they face the same 1-to-many issues. Capturing that industry knowledge could help to improve the quality of work in many fields.
  • Cloning a family member: Kind of like Jor-El from superman, we could work together to create a shadow of a loved one through their messages, videos, audios, etc. Even if it’s not them, I know I would love to have met the grandmother I didn’t.

More Data Require: We need to capture more information on the existing catalogue and on more personas.

Improve Emulation Quality: In preliminary research, we learnt that RAG can only take you so far. Fine-tuning models is a better pathway to go down. A better production robust approach would be using QLoRA with a model like Gemma E2B/E4B.

Use Sovereign Versions: We relied on ElevenLabs for voice cloning, for example, maybe a local version with other pre-existing voice tech would reduce cost.

Production readiness:

The system currently lacks Auth and storage scalability via s3. Things like metadata structure, Zod, serverless functions on Vercel were used though.

If the judges want the deployment link, we can provide it, it uses APIs so we'd rather not publish publicly on the site right now to avoid cost drain from other competitors. We can swap the keys for free allocation versions post-competition if historical showcasing is needed.

Built With

Share this project:

Updates