.png&w=256&q=75)
🤓 Latest Submissions
HangON: Voice Front Desk for Solo Trades
Solo tradespeople lose jobs every day because their hands are busy when the phone rings. A caller with a leak won't leave a voicemail; they call the next number. HangON is a voice front desk that answers for them. The call runs on AssemblyAI's Voice Agent API: speech to text, the LLM and the voice over a single WebSocket, with barge-in and adaptive turn detection. We use balanced mode so callers who pause mid-thought ("Tuesday... actually, no, Wednesday at 2") can finish before the agent replies. The agent acts through client-side tool calls. check_calendar_availability tests the caller's requested time against real booked jobs and business hours and returns genuine open slots if it is taken. book_service_appointment saves the job to Postgres only after the agent reads it back and the caller says yes. send_confirmation_email sends the confirmation through Resend from our own domain, after the agent asks for the address and confirms the spelling. AssemblyAI's LLM Gateway turns the live transcript into structured job fields with a strict JSON schema; anything the caller never said stays empty instead of being guessed. After the call, the same gateway writes the technician a brief: the problem, the caller's mood, and parts to bring. Our server mints a short-lived voice token for each call, so API keys never reach the browser. HangON is built to be honest: no placeholder data, no false success messages, and clear errors instead of canned answers.
30 Sep 2026
REGRET: Autonomous AI Options Trading Agent
REGRET is an autonomous quantitative options trading agent built for the Alpaca AI Trading Agents Hackathon. It addresses the fundamental flaw of traditional AI trading bots, specifically hallucinated entries and unbounded tail risk, by strictly separating open source probabilistic reasoning from deterministic code execution. 1. Strategic Intelligence Layer: The system continuously screens liquid underlyings (SPY, QQQ, NVDA, IWM) for elevated Implied Volatility regimes where IV Rank exceeds 50%. Using Featherless.ai serverless inference with open source quantitative models (Qwen 2.5 72B and Llama 3.3 70B), REGRET analyzes macroeconomic context, volatility skew, and support and resistance levels to formulate defined risk credit spread setups (Bull Put Spreads and Bear Call Spreads). 2. Six Deterministic Python Risk Gates: Before any order reaches the broker, it must pass all six hardcoded mathematical risk gates: (1) Max Loss Cap: Hard ceiling at $500 max loss per trade with zero naked risk. (2) Daily Drawdown Halt: Trading halts if cumulative daily loss exceeds $2,000. (3) Position Ceiling: Maximum 5 concurrent hedged spreads. (4) Liquidity and Spread Health: Rejects contracts exceeding a 10% bid ask spread to credit ratio. (5) Greeks Validation: Requires positive net Theta and bounded Delta between minus 0.40 and plus 0.40. (6) Expiration Safety: Limits expiries between 7 and 45 DTE to prevent gamma explosions. 3. Alpaca Execution and Lifecycle Management: Approved trades are routed natively through the Alpaca multileg options API. An active position manager continuously monitors the book, executing automated exits at 50% profit target, 2.0x initial credit stop loss, or 1 DTE expiration guard to eliminate assignment and pin risk. REGRET provides full CLI tooling, MCP server endpoints, and an interactive real time dashboard.
4 Sep 2026
