How SeventyFly AI Automates Dota 2 Match Analysis from Data to Results

Oleh Komarnitsky — Tue, 26 May 2026 15:15:33 +0000

A technical breakdown of the automation-first pipeline behind SeventyFly AI: match discovery, data refresh, prediction output, result validation and public statistics.

TL;DR

SeventyFly AI is an automated esports analytics platform built around Dota 2 match data, machine learning and real-time synchronization.
Dota 2 analytics is hard because heroes, patches, meta shifts, team form and noisy data make static analysis unreliable.
The core workflow is automated end to end: match discovery → data refresh → prediction generation → public output → result validation → public statistics update.
Automation matters because manual prediction workflows do not scale and can introduce selection bias into the public record.
The public layer shows win probability, match context and live statistics. As of the current public 2026 record on the project website, SeventyFly AI tracks 59.86% winrate across 426 predictions.
Performance is validated by outcomes, not claims: winrate, profit in flats and ROI help show whether the model’s probability output works in real betting-market conditions over time.
The model stays private: users see clean, understandable outputs while internal scoring logic remains protected.
Predictions are probabilistic, not guarantees. The project is built around transparent tracking, not curated highlights.

Introduction: Esports Predictions Are Usually Too Manual

Most esports prediction content still works like an opinion channel.

Someone watches a few matches, checks recent form, reads community discussions and posts a confident call. Sometimes the analysis is good. Sometimes it is not. But the biggest issue is usually the same: there is no data pipeline, no automated result tracking and no public accountability after the match ends.

SeventyFly AI was built to be different.

The goal was not to create another Telegram channel where a person manually picks matches and writes instinct-based predictions. The goal was to build an automated Dota 2 analytics system where match data drives the process, every analysis uses fresh context and every tracked prediction eventually faces the actual result.

That sounds simple. Building it is not.

Dota 2 is one of the most difficult games in esports to analyze. There are more than 120 heroes, constant balance changes, unstable team form, different tournament levels and a competitive meta that can shift quickly after a patch. The data is also messy: team names can be inconsistent, match statuses change, results arrive with delays and historical patterns can become outdated faster than expected.

Building a serious analytics product in that environment requires more than a model. It requires a full system: data collection, normalization, freshness checks, automated publishing, result validation, public statistics and safe separation between what users see and what the model uses internally.

Main SeventyFly AI dashboard with match cards, live analytics and public statistics.

Why We Built SeventyFly AI

I started working on SeventyFly AI in October 2025. The project began as a technical attempt to move Dota 2 predictions away from subjective calls and closer to measurable, automated analytics. Public testing started in January 2026, with the goal of making the system accountable from the beginning. Since then, public statistics on the website have been updating in real time as new tracked predictions are settled.

The frustration was simple: in esports, a lot of prediction content looks confident, but very little of it is properly tracked.

Some channels only publish successful outcomes. Some never show long-term statistics. Some mix different types of analysis into one number. Others depend too much on human selection, which makes the record hard to trust.

We wanted to build something cleaner: a system where the core prediction lifecycle is automated, where statistics are generated from tracked outcomes and where the public can verify performance over time.

Automation became the founding engineering decision.

Manual match-by-match selection introduces bias. The moment a human decides which matches to skip, the public statistics start reflecting editorial judgment rather than the system's actual output. For SeventyFly AI, the human role is not to manually pick every match. The human role is to design, monitor and improve the infrastructure: data quality, product logic, user experience, safety and long-term system reliability.

The Core Idea: Automation-First Architecture

The central design principle of SeventyFly AI is that the prediction lifecycle should run without manual intervention on each match.

The simplified workflow looks like this:

Match Discovery → Data Refresh → Context Normalization → Prediction Generation → Public Output → Website / Telegram Sync → Result Detection → Prediction Validation → Public Statistics Update

Each step feeds the next.

When a Dota 2 match becomes available in the competitive data stream, the system picks it up, refreshes the relevant context and prepares the public output. The prediction is then synchronized with the website and Telegram ecosystem.

After the match ends, the result is connected back to the original prediction. The public statistics update from tracked outcomes rather than manual edits or curated highlights.

This is the core difference between an automated analytics product and a manual prediction channel.

There is no manual match selection for every prediction. There is no choosing when to publish based on how confident a person feels that day. There is no manual rewriting of the statistical record, except for clear technical errors, duplicate entries or invalid matches.

The prediction lifecycle itself belongs to the pipeline.

Automation-first prediction lifecycle behind SeventyFly AI — from match discovery to public result tracking.

The Product Layer: Dashboard and Telegram

SeventyFly AI is not just a backend model. It is a synchronized product ecosystem built around the automated pipeline.

The main public dashboard is seventyfly.com. The root domain opens the live analytics experience directly, so users land on the actual product instead of a static landing page. The dashboard shows active and upcoming match predictions, win probability outputs, team context and public statistics in one place.

The interface stays synchronized with the backend pipeline so users can check the current state without digging through raw data. The goal is to make the product useful at a glance: match cards, filters, public statistics and recent outcomes are all part of the same live surface.

Match cards are intentionally simple. Users see the information that matters for public understanding: teams, tournament context, win probability, match context and result status when available. Internal scoring details stay private.

Telegram is the distribution layer. The free channel publishes public predictions and informational analyses, while the private channel is designed as a cleaner feed for tracked predictions and users who want less noise.

Public statistics are generated automatically from tracked outcomes. As of the current public 2026 statistics shown on the project website, SeventyFly AI tracks 59.86% winrate, +47.15 profit in flats, +11.04% ROI and 426 predictions. These numbers are not manually selected for appearance. They update as the system connects predictions with real results.

Wider dashboard view with filters, match cards and public statistics.

Single match card showing teams, win probability and public match context.

Telegram prediction post format used as a distribution layer for SeventyFly AI.

The Data Pipeline

If you strip away the interface, SeventyFly AI is fundamentally a data pipeline.

Raw Dota 2 match data → Normalized entities → Match context assembly → Prediction output → Live display → Result detection → Performance statistics

The pipeline starts with professional Dota 2 match data from public and official data sources. Raw data is never as clean as people expect. Team names can differ between sources. Tournament structures are not always consistent. Hero entries need to be normalized. Match statuses change. Results can arrive late.

Every major entity has to be structured before it becomes useful: teams, heroes, tournaments, leagues, patches, match status indicators and historical records. Without clean entity resolution, the downstream analytics becomes unreliable.

Heroes are a good example. A public data source may provide hero IDs or raw draft entries, but the product needs those entries to become readable context: hero names, draft structure, match association and later result connection. The same is true for teams and tournaments. The system has to turn raw records into stable entities before they can be shown on a live page or connected to a historical result.

Freshness matters just as much as structure.

Before a match is shown to users, the system refreshes the relevant context so the public output is based on the latest available data rather than a static snapshot from hours earlier. This matters in Dota 2 because the picture can change quickly: roster substitutions, updated match status, tournament context, draft information and recent results can all affect how a match should be presented.

After a match is finished, the result is connected back to the original prediction so public statistics can be updated automatically. This step is not glamorous, but it is one of the most important parts of the product. Without result validation, predictions become disconnected records with no accountability.

In practice, this means the platform is not only generating a prediction before a match. It also has to remember what was published, detect the final outcome, resolve edge cases such as delayed or invalid matches, and update the public record consistently. That is the difference between a prediction script and a production analytics system.

The hardest part was not just generating a prediction. The harder part was building the system around it: data freshness, automation, result validation, public statistics and safe public display. Those layers are what turn a model into a real product.

Public statistics block showing the current tracked record: 426 predictions, 255 wins, 171 losses, 59.86% winrate, +47.15 flats and +11.04% ROI.

Heroes, Patches and Match Context

Dota 2 is not like traditional sports analytics.

The game changes at a fundamental level with every major patch. A hero that dominates one patch can become average or even irrelevant in the next. Team strategies built around specific hero combinations can lose value overnight. Historical hero win rates can shift dramatically after balance changes.

That creates a continuous challenge: how much historical data is still relevant? How much should recent context matter? How quickly should the system adapt when the meta changes?

The answer is not a single static rule. It requires tracking patch context alongside match data, understanding which heroes are currently relevant and being careful about how historical patterns are interpreted when the game state changes.

This is one of the reasons Dota 2 analytics is technically interesting. The problem is not only “which team is stronger?” It is also “which data still describes the current game?”

On the product side, match context helps users understand why a prediction exists instead of seeing a raw probability number with no explanation. On the internal side, the model logic remains private, but the public product still needs to communicate enough context to be useful.

Static snapshots are not enough. In esports, data can go stale faster than most people expect.

The AI and Model Layer

At the core of SeventyFly AI is a model that produces a probability-based view of a match.

The output users see is a win probability: a number between 0 and 100% expressing which team the model sees as having an edge based on the available context. It is not a certainty. It is not a guarantee. It is a probabilistic assessment of one likely outcome.

The model combines multiple categories of signals around teams, match context, historical data and current conditions. The exact scoring logic, weighting and internal model design are intentionally not public. That is the proprietary part of the project.

What matters publicly is the lifecycle around the model.

Predictions that match the public tracking rules are included in the performance record automatically. The goal is to keep tracked statistics consistent and avoid mixing different types of public outputs into one misleading number. Analyses outside the tracked scope can still have informational value, but they are kept separate from the official performance record.

The system is designed for continuous evaluation. As new matches come in, predictions are tracked, results are connected and the public record remains visible.

Some predictions are wrong. That is expected. A probabilistic system should not be judged by one match, but by meaningful sample sizes over time.

Predictions are probabilistic assessments, not financial advice. Past performance does not guarantee future results.

Public Transparency and Result Tracking

One of the most important product decisions was to make result tracking part of the system from the beginning.

A product that only highlights wins is easy to build. A product that tracks everything — including losses — requires treating result validation as core infrastructure.

Every tracked prediction has a lifecycle: it is generated, published, matched to a result and included in public statistics. The system is designed so tracked predictions remain part of the record, except for clear technical errors, duplicate entries or invalid matches.

As of the current public 2026 record on the project website, SeventyFly AI shows:

59.86% winrate
+47.15 profit in flats
+11.04% ROI
426 predictions published

Winrate alone is not enough. A model can be right often and still fail if the market price is wrong, or be wrong sometimes and still perform well if the probabilities identify value over time. That is why SeventyFly AI also tracks profit in flats and ROI. These metrics connect the model’s public predictions to real betting-market conditions without claiming that any single match is guaranteed.

In that sense, betting results are not the identity of the project — they are the external performance test. If a probability model is useful, it should eventually show up in tracked outcomes: not only in a clean winrate number, but also in profit, ROI and long-term accountability.

The point is not to claim certainty. The point is to create a transparent public record where the system has to stand behind its own output.

Public accountability is a product feature.

Public Data Safety and Engineering Maturity

As the platform grew, another principle became clear: public product data and internal model data must be separated.

Early in an AI product, it can be tempting to expose too much. More fields can feel like more transparency. But raw internal data usually creates two problems: it confuses users who need a clean product experience, and it exposes proprietary signals that should remain private.

The solution is a clean public data layer.

The website and Telegram outputs receive fields selected for public value: win probability, team context, match metadata and result status. Internal scoring data flows through separate paths and does not reach public endpoints.

This is not only a security decision. It is a product decision.

Users do not need raw model internals to understand the platform. They need a clear output, reliable context, honest result tracking and a consistent public record.

That separation is part of building a serious AI product rather than a prototype.

What We Learned

Building SeventyFly AI produced lessons that apply far beyond Dota 2.

Esports data is noisy. Team names, match IDs, tournament structures and results are often inconsistent across sources. Entity resolution is not exciting, but it is where reliability begins.

Data freshness matters more than raw data volume. In a fast-changing game, a large amount of stale data can be less useful than a smaller amount of current context.

Automation is the product. A manual workflow can produce individual predictions. An automated pipeline produces a trackable, scalable and less biased record.

Result validation must be first-class infrastructure. If you do not connect predictions to outcomes automatically, public statistics become fragile and difficult to trust.

Public output and internal model logic should be separated early. Public fields need to be clean and understandable. Internal fields need to remain precise, complete and protected.

Predictions should be treated probabilistically. A 65% win probability prediction can still lose. That does not automatically make it a bad prediction. The only honest evaluation happens over meaningful sample sizes.

Product UX matters as much as model quality. If users cannot understand the output, they cannot benefit from the analytics.

What's Next

SeventyFly AI is already a working automated system, but there is still a lot to build.

The next direction is deeper live analytics: moving beyond pre-match outputs toward richer real-time context as more live data becomes available.

Public statistics can also become more detailed: breakdowns by hero pool, tournament level, meta period and other public categories would make the record more useful for users who want to understand performance beyond one headline number.

Another major direction is user infrastructure: Telegram authentication, user accounts, subscription status and a cleaner personal cabinet for users who follow the project regularly.

Longer term, SeventyFly AI can become more conversational. An AI Agent connected to the project’s own data could let users ask about teams, heroes, match history, public statistics and Dota 2 trends without exposing the internal model logic.

The goal is the same as it was at the start: build a serious AI esports analytics product around data, automation and transparent tracking.

Explore SeventyFly AI

Main dashboard: seventyfly.com
About the project: seventyfly.com/about/
Free Telegram channel: t.me/seventyfly_free
Main Telegram channel: t.me/seventyfly_tool

If you are working on esports analytics, AI products or data engineering in fast-moving domains, I am open to partnerships, collaborations and technical conversations around what we are building.

SeventyFly AI is an esports analytics project. All predictions are probabilistic assessments based on available data. Nothing in this article or on the platform should be treated as financial advice, betting advice or a guaranteed outcome. Past performance does not guarantee future results.

DEV Community: Oleh Komarnitsky