Thousands of blockchain projects launch each year. Most fail. Some survive. Fewer still create durable value.
The blockchain ecosystem is vast, diverse, and fragmented. DeFi, infrastructure, gaming, tokenised real world assets, identity, payments thousands of use cases across dozens of sectors. Tracking platforms list upward of 15,000 tokens. Perhaps 2,000 to 3,000 show meaningful onchain activity. Only 300 to 500 are institutionally relevant.
The Scale Problem
This raises the question: how can one begin to analyse, understand or make sense of this landscape?
The honest answer is that traditional approaches cannot. A research team can deeply cover 20, maybe 50 projects. Traditional ratings agencies do not understand the mechanics of liquidity pools, token unlock schedules, or governance votes. Onchain analytics platforms provide raw data, transactions, wallet flows, contract calls, but raw data is not assessment. Consider a fund evaluating 50 DeFi lending protocols. A single analyst would need to continuously track onchain usage, token unlocks, governance votes, audit status, developer commits, and liquidity depth across all 50. That is not feasible. Now multiply that by six sectors and you begin to see the scale of the problem. There is no structured, scalable intelligence layer for this market. It does not exist.
First and core to this initiative, there must exist a standardised data layer that interfaces to the whole ecosystem. This is the foundation upon which everything else stands. If the data layer is wrong, nothing downstream works.
Why Taxonomy Matters
A standardised data layer necessitates an understanding of the structure of data that underlies the blockchain ecosystem. This understanding does not come from algorithms. It comes from subject matter expertise. A blockchain expert will map the data points that touch on the ecosystem: token economics, onchain usage, developer activity, governance structures, liquidity dynamics, social narratives, competitive positioning, compliance and risk posture. Each of these constitutes a distinct analytical domain. Together they form the vocabulary through which the system observes and describes any project.
The question then becomes: can we build an abstraction layer that categorises data into independent, non-redundant classes that together span the full ecosystem? To borrow from mathematical parlance: can we devise an orthogonal basis for blockchain project assessment? That is the design goal. Not perfect orthogonality in the strict mathematical sense, but domains sufficiently independent and collectively complete to describe any project in the ecosystem. This taxonomy, this basis, is the intellectual foundation of the venture. It is designed by humans, informed by deep domain knowledge, and it determines what the system can and cannot see. Get the basis right and the system can describe any project in the ecosystem. Get it wrong and there are blind spots that no amount of computation can compensate for.
Where Machine Learning Wins
With the data layer and taxonomy in place, machine learning enters. Machine learning is about learning from examples through statistical pattern recognition. Raw signals are engineered into features: measurable, computable attributes derived through domain informed transformations. A machine learning model then learns which combinations of these features are associated with observable outcomes, building a statistical mapping from inputs to outputs.
We provide a machine learning algorithm with large volumes of these input-output mappings, project snapshots across time, across sectors, across lifecycle stages, and it builds a view of the ecosystem. It learns which feature combinations are associated with project health, and which are associated with deterioration or failure. Dependencies between inputs emerge. Patterns no human could track across hundreds of projects simultaneously become visible. A declining developer commit rate, combined with a concentrating token holder base, combined with a governance participation drop, tells a story invisible when each metric is viewed in isolation.
The ability of ML to then generalise to new unseen data is key. A trained system, when presented with a project it has never evaluated before, produces an assessment grounded in everything it has learned from every project it has seen. Therein lies the power of ML and the reason it is the only viable approach at this scale.
It is worth noting that the methodologies underpinning modern machine learning are not new. Most have existed since the late eighties, when backpropagation and neural networks emerged. What changed is compute. Cheap, cloud-based computing has enabled the practical implementation of these methods at scale, which in turn has created a feedback loop driving further advances. The tools are mature. The application to this domain is what is new.
Prediction, Confidence and Explainability
Such a framework does not produce point estimates. It provides outputs that quantify their own uncertainty, not single numbers. The system communicates not only what it believes but how confident it is. An institutional investor needs to know not just that a project is flagged, but how much weight to place on that flag. A system that says "this project has a 73% likelihood of maintaining health over the next 60 days, with a confidence interval of 64% to 81%" is fundamentally more useful than one that says "this project is rated B+."
Through disciplined model design, hyperparameter optimisation and feature importance are built into the training and inference process. This means two things. First, we guard against overfitting, which is the tendency of a model to memorise its training examples rather than learn generalisable patterns. A model that merely memorises is useless in production. Second, the features that drive each output are identified and ranked. This adds interpretability and transparency to every result. We can say not only what the assessment is, but why, which features drove it, and by how much. Combined with a natural language layer that translates these attributions into prose, this enables a narrative to be produced, which is essential for institutional trust. An institution cannot act on a black box. It must explain its rationale to its own stakeholders, to its compliance team, to its investors.
Context is Everything
Raw statistical power is insufficient without context. Cross sectoral and lifecycle aware context make or break such a venture. A project in its first months after launch behaves differently to a mature protocol with established market share. Low developer activity in a seed stage project is normal; in a mature project, it is a warning sign. A token unlock of 10% of supply is routine six months post launch; for a mature protocol with declining usage, it is a liquidity event. The same data point, interpreted without lifecycle context, leads to the wrong conclusion. This boils down to understanding the taxonomy of a project and how it evolves, from concept to launch to growth to maturity, and sometimes to decline. The system should be designed to know where a project sits in that arc and to adjust its assessment accordingly.
The Real Moat: The Pipeline
AI provides a mechanism through which subject matter experts can scale their judgement and expertise across an entire ecosystem. Dependencies are discovered. Input-output mappings are learned, prior assumptions are updated continuously and asynchronously whenever a new data point presents itself. Data points can range from token prices to social media narratives to GitHub repository updates to governance votes to smart contract audit results. The breadth is the point.
Successful deployment hinges on choosing and deriving features that could potentially impact the status of a project or digital asset from a risk, sustainability, or success likelihood perspective. The usage of these features, the topology of the underlying models, and all the units and components above are housed in the AI pipeline.
A key aspect of the pipeline is its ability to dynamically discover and adapt features, and to update its models in light of incoming data and changing market conditions. As new data arrives, models retrain continuously, incorporating the latest observations into their understanding of the ecosystem. Today's view is always more informed than yesterday's.
The pipeline is not merely a production system. It is a closed loop. Subject matter expertise is encoded into machine-readable formats such as scenarios, rules, and thresholds. The system tests these continuously against incoming data. But the loop runs in both directions. The system also surfaces dependencies and relationships that were not hypothesised in advance. These discoveries feed back into the design, sharpening the models, refining the taxonomy, and generating new hypotheses to test. In effect, a small research and development operation runs continuously in the background, improving the system with every data point processed. The pipeline executes and learns.
The pipeline allows more data categories, more feature definitions, the injection of subject matter expertise into assessment, and continuous model refinement as data accumulates. It is this pipeline not any single model or any single data source that enables scale. And scale is the only way to meet the opportunity.
Conclusion
Machine learning does not replace judgement. It scales it. The taxonomy defines what the system can see. The pipeline ensures it keeps seeing, learning, and improving as the ecosystem evolves. At this scale, with this complexity, across this many projects and sectors, there is no other way.