Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/alphaleaks60-maker/solvedocs2/llms.txt

Use this file to discover all available pages before exploring further.

Alpha Leak runs as a single process hosting 30+ concurrent background services, coordinated through a database and cache layer. Events are handled inline as they arrive — there is no queue between ingestion and processing, which keeps latency minimal and the operational footprint simple. The system is divided into four named phases. Each phase builds on the data produced by the one before it.

​
Pipeline overview

On-chain Events (real time)
        │
        ▼
┌───────────────────┐
│   EventProcessor  │  Decodes raw events into typed signals.
│                   │  Fan-out: routes each event to relevant services.
└────────┬──────────┘
         │  trade | create | complete
         │
         ├──────────────────────────────────────────────────┐
         │                                                  │
         ▼                                                  ▼
┌─────────────────────┐                        ┌─────────────────────┐
│   CORE SERVICES     │                        │   GENESIS WATCHER   │
│                     │                        │                     │
│  CurveManager       │                        │  60s observation    │
│  TradeWriter        │                        │  75-feature vector  │
│  VelocityTracker    │                        │  ONNX inference     │
│  CandleAggregator   │                        │  genesis_signal →   │
│  WalletDiscovery    │                        │  signal channel     │
│  WalletEnrichment   │                        └─────────────────────┘
│  OutcomeTracker     │
│  CreatorTracker     │
│  BotDetector        │
│  RiskScorer         │
│  DataArchiver       │
│  MaintenanceService │
└─────────┬───────────┘
          │
          ▼
┌─────────────────────────┐
│   PHASE 1 — SCORING     │
│                         │
│  PeakTracker            │  Tracks 1h/4h/24h price peaks per signal.
│  PnlCalculator          │  Realised PnL, win rate, hold times.
│  FeatureComputer        │  30-min feature roll-up per active wallet.
│  WalletScorer           │  7-component alpha score (0–100).
└─────────┬───────────────┘
          │
          ▼
┌─────────────────────────┐
│  PHASE 2 — INTELLIGENCE │
│                         │
│  SignalScorer           │  Composite signal quality score.
│  TokenLifecycle         │  8-state lifecycle classifier (60s cadence).
│  BundleDetector         │  5s time-bucket cluster detection.
│  CreatorRiskScorer      │  Rug rate, insider pct, serial detection.
│  CoOccurrence           │  Wallet pair co-buy frequency graph.
│  GraphBuilder           │  Graph analysis (hourly).
└─────────┬───────────────┘
          │
          ▼
┌─────────────────────────────────────┐
│  PHASE 3 — ML & ADVANCED INTEL      │
│                                     │
│  MlInference         (5s cadence)   │  68-feature ONNX scoring.
│  AntiSignalEmitter   (30s cadence)  │  Multi-trigger adversarial detection.
│  DeepDiveWorker                     │  Enriched wallet analysis.
│  CopyTradeDetector   (15m cadence)  │  3-type copy-trade classification.
│  OperatorDetector                   │  Coordinated operator identification.
│  AlphaDecayTracker   (60m cadence)  │  Per-wallet signal decay curves.
│  SignalCrowdingDetector (60s)       │  Crowding ratio → score penalty.
│  MarketRegimeDetector (10m)         │  4-state regime classification.
│  ModelMonitor                       │  Live model drift detection.
└─────────┬───────────────────────────┘
          │
          │  signal channel (real time)
          ▼
┌─────────────────────────┐
│     LIVE TRADER         │
│                         │
│  InlineScorer           │  ONNX inference inline (no DB round-trip).
│  PortfolioManager       │  Phase-gated positions + circuit breakers.
│  TradeExecutor          │  Bonding curve execution, atomic bundles.
│  ExitMonitor            │  3s poll: TP / SL / timeout / anti-signal.
└─────────────────────────┘

​
The four pipeline phases

1

Phase 1 — Scoring

Scoring runs on a continuous 30-minute cycle across all active wallets. PeakTracker records price peaks at 1h, 4h, and 24h horizons for every signal. PnlCalculator computes realised PnL, win rate, and hold time distributions. FeatureComputer rolls up a per-wallet feature set, and WalletScorer distils that into a 7-component alpha score between 0 and 100.
2

Phase 2 — Intelligence

Intelligence services build structured knowledge on top of raw event data. TokenLifecycle classifies each active token across 8 lifecycle states every 60 seconds. BundleDetector clusters trades into 5-second time buckets and flags coordinated entries. CreatorRiskScorer tracks rug rate, insider percentage, and serial-rugger patterns. CoOccurrence and GraphBuilder maintain a wallet co-buy frequency graph, updated hourly.
3

Phase 3 — ML and advanced intelligence

ML inference runs every 5 seconds, scoring all active tokens against a 68-feature ONNX model. AntiSignalEmitter scans every 30 seconds for tokens that have triggered two or more adversarial rules simultaneously. CopyTradeDetector runs every 15 minutes to classify wallet behaviour into three copy-trade types. MarketRegimeDetector reclassifies the global market state every 10 minutes across 4 regimes. AlphaDecayTracker maintains per-wallet decay curves across 8 delay buckets, updated hourly.
4

Live trader

The live trader subscribes to the real-time signal channel and acts on signals from all three emitters — SignalEmitter, AntiSignalEmitter, and GenesisWatcher. InlineScorer runs ONNX inference without a database round-trip. PortfolioManager applies phase-gated position limits and circuit breakers. TradeExecutor submits atomic bundles to the Pump.fun bonding curve. ExitMonitor polls every 3 seconds and triggers exits on take-profit, stop-loss, timeout, or anti-signal.

​
Data stores

StoreRole
DatabasePrimary store for all trades, signals, wallet profiles, ML scores, detected bundles, copy-trade pairs, regime snapshots, and live trade history. Trades table is monthly-partitioned for query performance.
Cache layerReal-time pub/sub channel for signal fan-out to the live trader; key-value cache for wallet stats, crowding ratios, market regime, and bonding curve state. Also maintains the tracked wallet and known bot sets.
Long-term archiveRaw trade data older than 60 days is automatically archived for storage efficiency.
ONNX model filesLoaded at startup and hot-reloaded every 5 minutes. Stored alongside metadata files describing features, calibration parameters, and PR-AUC.

​
Concurrency model

Every background service follows the same pattern: a timed loop that checks a running flag before each cycle, skipping if the previous run hasn’t finished. This prevents overlapping database queries under load without requiring a separate job queue. Memory usage is monitored and logged after every run.

​
Inter-service communication

Services communicate in two ways: Database reads — most intelligence services read each other’s output from the database. For example, MlInference reads WalletScorer alpha scores; AntiSignalEmitter reads BundleDetector results. Signal channel — the real-time bus for the live trader. SignalEmitter, AntiSignalEmitter, and GenesisWatcher all publish to it. The live trader subscribes to all three types. The API server streams the channel to connected clients via Server-Sent Events.

Build docs developers (and LLMs) love

Get started for freeTalk to us