Bittensor Subnet 15: BitQuant

What can BitQuant actually do? It enables a wide range of quantitative analyses and decisions in DeFi via a simple question-answer interface. According to OpenGradient (the team behind BitQuant), the subnet “implements a decentralized AI framework for quantitative DeFi analysis”. Through natural language queries, users can tap into:

• ML-powered market analytics – e.g. analyzing price trends, on-chain metrics, or market sentiment
• Portfolio analysis and optimization – e.g. assessing your wallet’s asset allocation, volatility, or suggesting rebalancing strategies
• Risk assessment and trend analysis – e.g. evaluating liquidation risks, drawdowns, or identifying emerging trends in DeFi
• Quantitative strategy evaluation – e.g. backtesting or comparing yield strategies across protocols

These are precisely the kinds of advanced analytics that were previously limited to quant firms or heavy bespoke tools. BitQuant’s high-level mission is to “democratize access to advanced quantitative analytics for DeFi and crypto markets” and “enable composable, on-chain quantitative strategies and risk metrics” by crowdsourcing many AI agents into one network. The subnet incentivizes high-quality AI agents to provide real-time, actionable insights, fostering a robust ecosystem of decentralized, AI-driven financial intelligence. In essence, BitQuant wants to bring hedge-fund-grade analysis to anyone’s fingertips, within seconds, via an AI agent that understands both natural language and the complexities of decentralized finance.

Notably, BitQuant has been battle-tested with real users. Prior to its official subnet launch, the BitQuant engine went through months of beta testing and stress-tests. The team reports that in the lead-up to launch, BitQuant handled 4.7 million sessions, exchanging over 41 million messages with 1.5 million unique users, with an average session lasting 4 minutes. These weren’t toy examples; users asked real, high-stakes questions like the SOL depeg scenario mentioned above, or “What’s my wallet’s current liquidation risk?”, etc., and the system learned and improved from each interaction. This proven usage shows that BitQuant can scale to handle large demand and complex queries, sharpening its protocols under real market stress before going fully live on Bittensor.

Overall, BitQuant (Subnet-15) acts as a decentralized “quantitative analyst” for crypto, available 24/7. It reacts to market events and reasons about data in the same block, aiming to deliver the best of both speed and depth of analysis. In the words of its creators, it’s “DeFi’s first open market for on-chain intelligence”, where AI agents can “react and reason – in real time, on-chain” to answer the questions traders care about. By combining fast reaction with sound reasoning – and backing it all with on-chain verification – BitQuant provides a novel way for traders to get trustworthy insights on the fly, directly from a network of AI experts rather than a single centralized service.

BitQuant’s product is two-fold: it is both an AI agent framework for DeFi analysis and a Bittensor subnet deployment of that framework. In practical terms, OpenGradient (the developer) built a full software stack (BitQuant framework) that can parse natural language questions about crypto, fetch on-chain/off-chain data, and produce answers – and they have deployed this stack as Subnet-15 on the Bittensor blockchain, so that it runs in a decentralized, incentivized network of nodes. The “consumer-facing product” is accessible via a simple UI (e.g. the BitQuant web app at bitquant.io) or API where a user asks a question and gets an answer. But behind the scenes, that query is handled by BitQuant’s distributed infrastructure: a set of miner nodes and validator nodes running on Bittensor, all executing the BitQuant AI logic. The result is an open-source, on-chain service that any developer can also integrate into their own dApp via an API, or even fork to create their own similar subnet.

From a technical architecture perspective, BitQuant is designed in a modular, layered fashion. According to OpenGradient’s documentation, the BitQuant system has “three layers” in its data path:

Client/API Layer – This is the user interface or integration point. It could be a CLI tool, a chatbot, a web app, or any application that sends queries. BitQuant provides a simple API endpoint (POST /agent/run) via a FastAPI server for external apps to query the agent. Developers can thus plug BitQuant into dashboards, wallets, or trading bots with a single import using the provided QuantAPI client library.

AI Decision Layer – This is the “brain” that decides which expert agent should handle a given query. BitQuant uses a Router LLM (a large language model) that interprets the incoming question and routes it to the appropriate specialist agent. For instance, BitQuant currently has (at least) two main expert modules: an Analytics Agent (for things like market trends, metrics, risk analysis) and an Investment Agent (for actionable trade insights and execution logic). The router will choose the Analytics agent for a question about, say, TVL trends, versus send a question about executing a trade to the Investment agent. This design makes the system agentic: different sub-agents focus on different domains, and the router LLM picks the right tool for the job. (Currently the Investment Agent is Solana-focused for trade execution, but it’s built to add EVM support next.)

Data & Execution Layer – This is where the chosen agent actually fetches data and/or performs on-chain actions to construct an answer. Each agent in BitQuant comes with connectors to various data sources and protocols. For example, agents can directly query on-chain data from Solana’s RPC, pull market info from CoinGecko, get DeFi analytics from DeFiLlama, or even interact with DeFi protocols like Orca, Solend, Kamino (Solana yield vaults), etc.. This allows the agent to gather real-time prices, liquidity info, yield rates, etc., and even simulate or execute transactions (e.g. the Investment agent can interface with a DEX or vault on Solana). The results are then compiled into a response. Notably, BitQuant’s agents produce not just an answer in text, but also a verifiable proof or signed output (the “synapse”) that can be confirmed on-chain. Because all the data fetching and computation is done in a deterministic way (or with cryptographic signing), anyone can later audit that the answer was derived from real data and correct logic.

This entire architecture is packaged as an open-source codebase. The BitQuant repository (now open under MIT License) provides all the components: the QuantSynapse schema that defines the standard format for questions and answers, the agent logic, and templates for running Bittensor nodes. Impressively, the core definitions are very lightweight – “QuantSynapse defines the request/response schema in just 35 lines” of code. Similarly, the core logic for miners and validators is only a few hundred lines (around 300 each), demonstrating the minimalist, composable design. By keeping the protocol simple and open, anyone can audit it easily, extend it, or even “spin up your own subnet tomorrow” using the same framework. The current build includes a Python client library (QuantAPI) to query the subnet, and reference implementations for a miner and validator node – all intended to lower the barrier for developers to participate or build atop BitQuant.

The BitQuant subnet on Bittensor consists of two types of nodes: Miners (servers running the AI agents) and Validators. Both are necessary parts of the product’s decentralized infrastructure:

Miner Nodes (BitQuant AI Agents) – Miners are the workhorses that handle incoming queries. When a miner receives a validated query, it will run it through the BitQuant agent logic: first performing any blacklist/priority checks, then having the appropriate internal agent (analytics or investment module, etc.) process the query. The miner gathers the required data (via the APIs mentioned above), computes the answer, and then packages the response along with metadata like confidence or proof. Miners basically execute the AI analysis and produce the Q&A result. They also handle errors or edge cases (for example, if data sources are unavailable, the miner must handle that). Because running the BitQuant agent (especially the large language model routing and multi-source data fetching) is computationally intensive, miners require significant hardware – the team recommends GPUs and a solid CPU, as outlined in their hardware specs (the typical setup includes an 8-core CPU, 32GB RAM, and ideally a GPU for the LLM). Anyone with the required hardware can run a miner node and start contributing answers to the network.

Validator Nodes – Validators are the coordinators and quality controllers of the subnet. A validator’s job is to sample and evaluate the miners. When a user question comes into the network (likely reaching a validator first via the API or a gateway), the validator will select one or more miner nodes (often randomly or based on some scheduling) and forward the query to them. Once the miners respond with their answers, the validator scores the responses using a predefined reward function and assigns a weight or “score” to each miner based on the accuracy and value of their answer. If multiple miners answered, the validator can compare them or aggregate the results to determine which was most accurate. These scores are then used to update the metagraph (the Bittensor network’s record of each miner’s performance metrics). Validators thus control the reward distribution: miners that provide better answers (higher accuracy, faster response, etc.) will gain weight and earn more “alpha” token rewards, whereas miners that give poor answers will see their weight decay. Validators can even initiate slashing (penalizing a miner’s stake) if a miner consistently performs badly or maliciously, as noted earlier. In summary, validators keep the miners honest and the network healthy, ensuring that high-quality intelligence is being provided by the subnet.

Put together, the product is essentially a decentralized Q&A system for DeFi analytics, implemented as a Bittensor subnet. The OpenGradient team has delivered everything needed to run this system: the AI agent software, the blockchain integration, and even a simple frontend. Users can interact with BitQuant by visiting the BitQuant web app or via API calls, ask natural language questions about crypto, and receive answers with the backing of on-chain verification. Meanwhile, under the hood, a network of independent operators (miners) are running the AI agents to supply those answers, and validators are measuring their performance and rewarding them in TAO (through the subnet’s α (alpha) tokens which are convertible to $TAO). Because it’s all open-source, others can build on top of BitQuant as well – for instance, integrating BitQuant’s capabilities into other trading platforms or wallets. In fact, OpenGradient encourages developers to extend the ecosystem: they envision custom agents (for things like NFT analytics, cross-chain comparisons, new strategies) being built and added, since “agents are modular and composable — route them, extend them, remix them”. The BitQuant framework is thus not just a single product, but a foundation for an evolving suite of AI-driven DeFi tools, all interoperating on the Bittensor network.