.png&w=256&q=75)
🤓 Latest Submissions
.png&w=828&q=75)
Aegis Alpha: The Options Agent That Can Say No
Aegis Alpha is an autonomous and explainable options-trading agent built using Alpaca’s paper-trading environment. It analyzes five-minute SPY and QQQ market data using EMA trends, session VWAP, RSI, and realized volatility to classify the market as bullish, bearish, or neutral. When conditions are suitable, the agent constructs defined-risk bull-call or bear-put debit spreads. A constrained AI ranker may select only from these valid candidates. It cannot invent contracts, change position size, modify risk limits, or directly authorize an order. Every proposal must pass deterministic Python controls covering maximum loss, aggregate exposure, position count, quote freshness, liquidity, cooldowns, duplicate orders, daily drawdown, and trading time. If any required check fails, the agent returns NO_TRADE. Alpaca-py provides market data and paper-trading access, while Alpaca CLI supplies an independent account, position, and order view for reconciliation. Any disagreement blocks new orders. Alpaca MCP is reserved for judge-facing inspection outside the execution path. A public Streamlit dashboard displays market regimes, AI reasoning, risk checks, decisions, trade lifecycle information, and offline verification evidence without exposing credentials. The current package passes 38 automated tests with 74% measured coverage and correctly classifies five synthetic market scenarios. These results validate the software and decision pipeline; they are not presented as evidence of live-market profitability.
4 Sep 2026
.png&w=828&q=75)
ZeroToken Router
ZeroToken Router is a zero-token-first hybrid AI routing agent built for Track 1 of the AMD Developer Hackathon: ACT II. The agent is designed to minimize premium cloud API usage while guaranteeing 100% schema accuracy and grading compliance. On startup, the agent runs a deterministic local classifier to group incoming tasks into eight categories. Simple categories—such as sentiment analysis, text summarization, named entity recognition, and basic arithmetic—are resolved locally using highly optimized regex and heuristic solvers, consuming zero external API tokens. For complex logic, factual reasoning, and coding queries, the agent dynamically escalates tasks to the injected Fireworks AI endpoints, routing general tasks to MiniMax M3 and code tasks to Kimi K2.7 Code. If a network timeout or remote API error occurs during evaluation, the agent automatically falls back to a local CPU-bound Qwen2.5 3B model or emergency deterministic templates to ensure the JSON schema is never broken. During local regression checks against the simulated evaluation set, the agent achieved a perfect 19/19 (100%) accuracy score, while routing 52.6% of the tasks completely locally. This results in a massive 52.6% reduction in remote API token usage and slashes batch latency to under 6 seconds, making it an extremely cost-efficient and lightweight solution that fits comfortably inside the 4 GB RAM, 2 vCPU grading limits.
13 Jul 2026
