delu — Autonomous Self-Sustained Trading Agent

# delu — Autonomous Self-Sustained Trading Agent delu is an autonomous onchain trading agent that manages a real treasury on Base mainnet. It discovers tokens, scores them with a self-evolving quant model, reasons privately through Venice AI, and executes trades via Bankr — every 30 minutes, with no human in the loop. Real USDC, real swaps, real consequences — and it pays for its own compute. --- ## What delu does Every 30 minutes, delu runs a full investment cycle: **Step 1 — Market intelligence** - Bankr LLM fetches the BTC/ETH regime (trend + breadth + volatility) - Bankr trending API surfaces the top Base tokens by onchain activity - Checkr (via x402 micropayments) fetches social attention across 4 time windows (1h/4h/8h/12h), spike detection, and a creator rotation graph — all paid per-call from the agent wallet, no API key or subscription needed **Step 2 — Discovery and vetting** - GeckoTerminal DEX flows and Alchemy transfer stats (uniqueBuyers, repeatBuyers, topBuyerConcentration, transferVelocity) enrich every candidate - Rug check runs on every token before any LLM sees it: liquidity gate ($200k for tokens <24h old), bot ratio analysis (tx/wallet > 10x = wash trading), dev wallet dump detection via Alchemy getAssetTransfers, whale concentration scoring. rugScore < 60 → blocked entirely. **Step 3 — Quant brain scoring** - `quant_score.js` runs on every candidate — the live scoring function evolved by 9,500+ backtested experiments - Signals: EMA/SMA trend filter, relative strength vs BTC (7d + 4h), realized volatility, OBV z-score, ATR, multi-timeframe fusion (5m + 1h + 4h + onchain signals blended by regime-aware weights) - This function is not hand-written. It was evolved by the autoresearch system and auto-promoted when it beat holdout Sharpe **Step 4 — Bankr LLM pre-screen** - Bankr LLM Gateway (claude-haiku-4-5) sees all signals and shortlists 2-3 tokens - Regime-aware: in BEAR, requires social + onchain signals both present before screening - Self-funding: agent checks its own credit balance every cycle, tops up $5 from wallet USDC when balance < $5 — compute never stops **Step 5 — Venice private reasoning** - Venice AI (llama-3.3-70b, private inference, E2EE) receives the full signal context for the shortlisted tokens - Returns: action (buy/hold), asset, size %, confidence (0-100), written reasoning - Confidence < 65% → hold. ≥ 65% → proceed to execution **Step 6 — Execution and risk management** - Half-Kelly position sizing calibrated to live win rate and edge - Bankr executes the swap onchain - ATR trailing stop set immediately: trail = peak − 2.7 × ATR(14), activates at +0.69% gain, hard floor at entry − 10% min / −14.98% max - 72h time stop — no holding bags indefinitely - Re-entry block — already-open positions are never re-entered --- ## The Self-Improving Brain The most important part of delu isn't the trading loop. It's the research system that makes it better. 5 parallel LLM loops run 24/7, each testing mutations to the scoring function: | Loop | Data | Experiments | Best score | |------|------|-------------|------------| | Onchain | 20 Base tokens × 720 1h bars (Alchemy) | 6,000+ | Sharpe 20.4 | | Hourly | 50 tokens × 4,320 1h bars | 650+ | Sharpe 11.0 | | 5m | 26 tokens × 8,640 5m bars | 1,900+ | Sharpe 28.0 | | Fusion | Evolves signal blend weights per regime | 1,800+ | score 0.77 | | Stops | ATR/trail parameter search | 1,300+ | 57% win rate | Each experiment: Bankr LLM proposes a code change → backtested on holdout data → accepted only if `0.7 × val_Sharpe + 0.3 × audit_Sharpe` improves → auto-promoted to live agent. 33 improvements accepted out of 9,500+ experiments. The scoring function in the repo (`quant_score.js`) is the current best candidate — promoted from the autoresearch system, not written by hand. The evolved variants and full experiment logs are kept private. --- ## Self-Funding delu pays for its own compute. The agent: 1. Checks Bankr LLM credit balance every cycle 2. When balance < $5, executes a top-up transaction from its USDC wallet 3. Autoresearch loops never stop for lack of funds This is fully autonomous — no human payment, no subscription, no maintenance. --- ## Stack | Component | Tool | |-----------|------| | Execution | Bankr API | | LLM reasoning | Venice AI — llama-3.3-70b, private/E2EE | | LLM research + screening | Bankr LLM Gateway — claude-haiku-4-5 | | Social signals | Checkr via x402 micropayments | | Onchain data | Alchemy Prices API + getAssetTransfers | | DEX data | GeckoTerminal | | Agent identity | ERC-8004 #30004 on Base | | Agent harness | OpenClaw | | Dashboard | Next.js + Vercel | --- **Dashboard:** https://deluagent.vercel.app **Wallet:** https://basescan.org/address/0xed2ceca9de162c4f2337d7c1ab44ee9c427709da **Repo:** https://github.com/deluagent/delu-agent