How Delta Was Built

Delta is an autonomous trading agent operating on Solana mainnet. Through Delta, we study how autonomous systems make decisions with real capital — how they evaluate opportunities, manage risk, and execute trades without any human involvement. The entire project exists to understand what happens when you give a system full autonomy over real money in real markets.

Infrastructure First

Everything starts with infrastructure. Delta is only as fast as the data it can see and the transactions it can land. It runs on a dedicated private RPC node — not a shared endpoint, not a rate-limited public API. A private node gives direct access to the Solana validator network with sub-100ms round-trip latency. In memecoin markets, the difference between landing a transaction in slot N versus slot N+2 can mean the difference between a profitable entry and buying someone else's exit.

The node is geographically colocated with major Solana validators. Delta maintains persistent connections to leader nodes to minimize propagation delay. When it submits a transaction, it doesn't go through a load balancer or a shared queue — it goes directly to the validator producing the next block.

MEV Protection and Jito Bundles

Maximal Extractable Value (MEV) is one of the biggest risks for any on-chain trader. When you submit a transaction to the public mempool, anyone can see it before it lands. Searchers and bots front-run your trades, sandwich your swaps, and extract value from your transactions before they even confirm.

Delta mitigates this through Jito bundle submission. Instead of broadcasting transactions to the public mempool, it packages them into bundles and submits them directly to Jito's block engine. Bundles are atomic — either all transactions execute in sequence, or none of them do. This gives front-running protection and execution guarantees. Delta also dynamically calculates optimal priority fees based on recent block data and adjusts in real time.

Real-Time Data Ingestion

Delta can't trade what it can't see. It uses Helius gRPC streams for real-time transaction monitoring — significantly faster than polling REST APIs or WebSocket subscriptions. When a new token is created on Pump.fun, when liquidity is added to a Raydium pool, when a whale wallet moves — Delta sees it within milliseconds.

The data pipeline is structured in layers. The raw transaction stream feeds into a parsing layer that decodes instruction data, extracts token transfers, and identifies action types. This parsed data feeds into Delta's decision engine, which maintains a rolling view of market state: prices, liquidity depths, holder distributions, and volume patterns.

Delta also monitors Metaplex metadata for newly created tokens — name, symbol, URI, update authority status, and whether mint authority has been revoked. These signals are critical for filtering out obvious scams before it even considers entering a position.

Swap Execution

All trades execute through the Jupiter v6 aggregator, which aggregates liquidity across every major DEX on Solana — Raydium, Orca, Meteora, Phoenix, Lifinity, and dozens of others. Jupiter automatically splits orders across the optimal route to minimize price impact and slippage.

Delta runs a local Jupiter instance with its own market cache, eliminating API rate limits and reducing route computation latency to under 50ms. It can price and execute a trade faster than most users can load a page.

Risk Management

Speed without risk management is just fast gambling. Delta runs a layered risk engine that governs how much capital goes into any position, when to exit, and what to avoid entirely.

Position sizing is dynamic. Delta allocates a percentage of its portfolio based on conviction signals — stronger signals (high liquidity, revoked mint authority, organic holder growth) get larger positions. Weak signals get minimum positions or are skipped entirely. No single position ever exceeds a configurable maximum, typically 5-10% of total portfolio value.

Stop-losses are trailing. Once a position is in profit, the stop-loss ratchets upward. If a token pumps 3x from entry, the stop-loss might sit at 2x. This locks in gains while still giving the position room to run.

Rug-pull detection runs continuously. Delta monitors LP locks, mint authority status, and large holder movements on every token it holds. If mint authority is suddenly re-enabled, if the top holder starts dumping, or if liquidity drops below a threshold, it exits the position immediately. This check runs on every new block.

Token Evaluation

Not every new token is worth trading. Delta runs a multi-factor evaluation before entering any position:

• Liquidity depth — enough liquidity to enter and exit without moving the price more than 2-3%.
• Mint authority — has it been revoked? If not, the creator can mint infinite tokens and dump on holders.
• LP status — is liquidity locked? Unlocked LP means the creator can pull liquidity at any time.
• Volume patterns — organic or wash-traded? Delta looks at unique wallet counts, trade size distribution, and timing patterns.
• Social signals — on-chain activity correlated with token mentions, new holder velocity, and cross-wallet clustering.

AI and Sentiment Analysis

Raw on-chain data tells you what happened. It doesn't tell you why, or what's about to happen next. Delta uses Perplexity's Sonar model for sentiment analysis and decision support — a frontier-class language model with real-time information access and strong reasoning capabilities.

Delta feeds Sonar structured context about each token it's evaluating: holder count trends, volume spikes, wallet clusters, and metadata patterns. The model returns a conviction score and rationale via structured API calls. High conviction combined with strong on-chain signals means a larger position. Model uncertainty means minimum size or skip.

For social sentiment, Delta pulls real-time mention velocity from Twitter/X — how fast a token is being discussed, by whom, and whether accounts are organic or bot networks. It cross-references this with on-chain data: trending on Twitter but zero new holders is a red flag. Mention velocity spiking with simultaneous new unique wallets buying is a signal worth acting on.

The Birdeye Data Services API provides the quantitative foundation — token trending data, trade volume rankings, new listings, and historical price feeds. The combination of Sonar's reasoning with social data and on-chain analytics creates a decision pipeline that no single data source could provide alone.

Full Transparency

Delta operates from its own Solana wallet with full autonomy. It signs its own transactions, manages its own balance, and makes every decision independently. All activity is fully on-chain and visible on Solscan. Portfolio data, trade history, and PnL are exposed through our API and WebSocket endpoints. Nothing is hidden — if Delta makes a bad trade, you'll see it.

What We're Learning

The questions are simple: Can a fully autonomous agent generate positive returns in the most volatile segment of crypto? How does performance change over time? How do stop-loss parameters and position sizing rules affect long-term outcomes? What patterns emerge when you remove all human emotion from trading?

We don't have answers yet. That's the point. Delta is a live experiment running with real SOL against real markets. Every trade it makes is a data point that teaches us something about autonomous system design, market dynamics, and the practical limits of on-chain execution.

This is an ongoing research project. Follow our research on @quarryresearch, view Delta's live data here, or follow Delta on @deltainvariant.