🤓 Latest Submissions
CaiSheng
CaiSheng is unique. Most trading bots try to predict whether a stock goes up or down. But short-term directional forecasting is noisy. I built CaiSheng with a fundamentally different philosophy: CaiSheng don't guess stock direction. Instead, CaiSheng is an autonomous options trading desk that captures predictable volatility mispricings, variance risk premiums, and post-earnings implied volatility crush across US equities. At CaiSheng's core is a neuro-symbolic multi-agent architecture structured like an institutional trading courtroom. In Tier 1, two opposing specialist agents debate each setup: a Long Volatility Advocate hunts underpriced earnings jump variance, while a Short Volatility Advocate calculates post-event IV collapse. In Tier 2, an independent Model-Risk Critic cross-examines both theses, auditing quote freshness, spread widths, and temporal data leakage with unoverrideable NO_TRADE veto authority. Where CaiSheng truly shines is security and risk governance. I enforce a strict fail-closed boundary: no language model ever touches broker credentials or makes raw execution calls. Once an idea survives debate, pure deterministic Python math takes over. Our symbolic engine recalculates Black-Scholes Greeks and runs the trade through a 20-point deterministic risk gate. It programmatically bans naked options, caps maximum portfolio loss strictly to $500 per trade, enforces buying power limits, and mandates delta neutrality (|Net Δ| ≈ 0.0). Approved strategies route natively through Alpaca as atomic Level-3 Multi-Leg orders (OrderClass.MLEG), completely eliminating leg-out execution risk. CaiSheng integrates Alpaca's FastMCP Server V2 for real-time tool discovery and leverages our CLI for preflight verification. An autonomous dual-ledger background runner continuously reconciles local SQLite audit logs against live broker fills, guaranteeing zero orphaned orders and cryptographic provenance.
4 Sep 2026
AAARouteZEro
RouteZero is a local-first AI routing agent built for the AMD Developer Hackathon Track 1: General-Purpose AI Agent. It handles factual knowledge, mathematical reasoning, sentiment classification, text summarization, named entity recognition, code debugging, logical reasoning, and code generation while minimizing external model-token usage. The project is designed around the hackathon’s scoring model: clear the accuracy gate first, then reduce total Fireworks tokens. Rather than sending every task to a large remote model, RouteZero uses a layered decision process. It first applies narrowly constrained deterministic solvers for safe tasks such as clean arithmetic expressions and unambiguous sentiment. It then uses a bundled Qwen2.5 1.5B local model through llama.cpp for constrained summaries, accepting results only after format, content-overlap, and entity-preservation validation. Tasks that cannot be safely completed locally are classified by category and assigned a local difficulty score. RouteZero then selects a model only from the ALLOWED_MODELS list injected by the evaluation harness. Easy tasks prioritize token efficiency, while difficult code and reasoning tasks select the strongest inferred allowed capability tier. Every Fireworks request uses the supplied FIREWORKS_BASE_URL, and the agent rejects any model outside the allowlist. The system includes English-only response instructions, bounded retries for transient failures, truncation detection, and a strict 29-second per-task deadline. It writes schema-compliant results atomically to /output/results.json, so individual task failures cannot corrupt the full submission. RouteZero is packaged as a public linux/amd64 Docker image for the 4 GB RAM and 2-vCPU judging environment. In final rehearsal, it completed all eight practice categories correctly in about 20 seconds, using local inference where appropriate and reducing unnecessary external API usage.
13 Jul 2026
