We are thrilled to announce ELNA - a new AI assistant built entirely on the Internet Computer blockchain that allows anyone to create customised chatbots tailored to their specific needs and interests.
With elna, you can upload your own data to train unique machine learning models that focus on niche topics and applications. Our goal is to make advanced conversational AI accessible for personal and professional use cases.
Connect/follow us on Twitter > https://twitter.com/elna_live
Why We Created elna
Current AI chatbots are limited in scope and constrained by the data their centralized operators choose to train them on. We envision an ecosystem of specialized AIs customized by users, for users — without being boxed in by a one-size-fits-all assistant.
By building elna on Internet Computer, we can deliver this vision in a transparent, permissionless, and decentralized way.
- Upload Documents & Datasets — Teach your AI specialist knowledge by providing domain-specific training data.
- Shape Personality — Give your chatbot a distinctive conversational style with customizable parameters.
- Deploy On-Chain — elna deploys your finished agent into secure canisters on Internet Computer.
- Dynamic Updating — Continuously expand your AI’s knowledge by uploading new data.
- Monitor Analytics — View usage metrics to see how your chatbot is performing.
Under the hood, elna’s capabilities are powered by:
- Canister smart contracts that handle model training and deployment.
- Vector database architecture for efficient on-chain data storage.
- Motoko components for access control, transactions, and analytics.
- A token-based usage model secured by the ICP protocol.
- Internet Identity integration for seamless user experience.
We welcome developers, researchers, and enthusiasts to get involved:
Visit : elna.live
- Try creating a sample chatbot with our sandbox environment (initial beta-testing waitlist is open)
- Provide feedback on potential features and improvements
- Help us spread the word about decentralized, customizable AI
Together we can pioneer the next evolution of AI - owned by communities instead of corporations. Join us on this journey!
All of my employees telephonic conversations are recorded. Can I upload my recordings and build the ultimate customer service AI voice bot?
Few questions -
How easily can ELNA be integrated into existing decentralized exchange platforms?
Is there documentation or an API available for integration purposes?
How can I train ELNA specifically to understand and handle cryptocurrency and DEX-related queries?
Looks awesome, Do you have a twitter?
Thank you for the interesting suggestion! Using real customer service call recordings to train a voice AI assistant is a great potential use case. We are excited that elna’s capabilities could be useful for building customized bots like this.
You’re absolutely right that being able to upload audio transcripts and leverage real conversational data would enable creating an incredibly powerful and natural customer service agent.
While our initial release focuses on text-based conversation models, expanding to voice is on our roadmap. We want to take the time to robustly develop the infrastructure for secure voice data ingestion and training.
We will add this to our feature requests and keep you posted on progress towards audio support. Please feel free to join our Discord where we discuss product development.
Thanks again for sharing this inspiring idea! Our goal is for elna to be flexible enough to fulfill many niche use cases like this customized voice assistant. We value your input on how our platform could empower new applications of AI. > Adding this as a priority feature to our product feature backlog
Hi @cakemaker1 , thanks for the great questions!
Regarding integrations, our goal is to make elna as seamless to incorporate into existing dapps and platforms as possible. We plan to provide thorough documentation and APIs to cover common integration use cases. [Now in beta ]
Join our wait-list
For decentralized exchanges specifically, we envision elna agents being useful as smart assistant bots to help users with queries related to trading, liquidity pools, and anything DeFi.
To train elna on crypto/DEX topics, you could upload relevant content like wiki entries, blog posts, whitepapers, support docs, and community conversations. The more niche data you can provide, the better elna will be at fieldering domain-specific questions.
You can also fine-tune her responses to use appropriate terminology and align with a DeFi product’s voice. We provide settings to customize tone, level of detail, etc.
Additionally, we will be integrating the training data as vector embeddings directly in the canister storage on Internet Computer. This provides more decentralization by keeping sensitive user data on chain rather than external servers.
I’d be happy to discuss further how we can tailor elna’s capabilities for DEX platforms. Please feel free to reach out directly on Twitter/X
feel free to connect as over twitter >> https://twitter.com/elna_live
thanks for the support
I joined the waiting list. Hopefully I can play around with it soon.
Also follow our Twitter https://twitter.com/home
for all the latest updates
What parameters can be configured? Which applications are configurable? Can I configure chatbots, Q&A, support bots, summarization, and text generation, among others?
Great question! Here are some details on the parameters and applications that can be configured with elna:
For any AI agent you build, elna allows customizing:
- Personality - Adjust tone, style, formality, voice
- Knowledge Source - Upload niche data to focus the training
- Response Details - Configure length, depth, and scope
- Interactivity - Set follow-up preferences and conversation flow
- Privacy - Manage data retention and visibility
These parameters can be applied to a wide range of agent types:
- Chatbots - For natural conversations and open dialog
- Q&A Bots - To handle common questions and searches
- Support Bots - For customer service and interactions
- Summarization - To process and condense long form content
- Text Generation - For creating original text like stories
We also leverage techniques like parameter efficient Instruction fine tuning to optimize models for specific tasks, beyond tweaking conversational settings. This adapts models to new domains efficiently.
Additionally, we can use zero shot and few shot prompting to tailor models for certain applications without extensive retraining. By providing just a few examples, we can specialize models for purposes like Q&A, search, content generation and more through prompting.
Between conversational parameters, efficient fine tuning, and prompting methods, elna provides multiple avenues to customize models for chatbots, personal assistants, content creation, customer service, and many other needs.
When are you planning to launch?
Thank you for providing access. I am currently working on customizing the exchange data and training it to meet my specific needs. I appreciate the opportunity to do this and look forward to returning in a week or two with some useful data. Thank you again for your assistance and support.
Have you started giving away whitelisting? I don’t see any response from your end for my request.
Absolutely excited to see ELNA’s innovative approach to AI assistants on the Internet Computer blockchain! This opens up a world of possibilities for tailored AI interactions. Your project resonates with my work on B3Wallet, a decentralized multi-chain wallet. We could potentially integrate an extension that provides signals to users for digital asset trading decisions. Looking forward to exploring collaboration opportunities and contributing to the evolution of AI. Kudos!
This is crazy and awesome.
How does the architecture work? Which model is running in a canister? I’m assuming there are still significant limitations in what you can do with LLMs on ICP.
I’d love to learn more about the architecture.
We are on closed alpha testing, Hopefully will be opening our beta release in coming month
Thanks for the interest @domwoe!
Our initial POC stores vector embeddings for user data in canisters while leveraging existing fine-tuned models externally. This allows creating specialized bots with custom knowledge.
Later we plan on running the inference engine fully on-chain using WebAssembly in canisters. Canister size is a limitation, but we think breakthroughs are achievable through techniques like quantization.
Now Training model on canister
We feel full fine-tuning may be overkill - but methods like LoRA for parameter efficient fine-tuning could adapt models efficiently.
Currently there is no plan to Pre training a model from scratch as its unnecessary and almost impossible in canister
Feel free to ask us any details on the same