.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
.png&w=828&q=75)
Options Alpha Agent
The agent runs unattended on a schedule, reads the volatility surface across eight liquid names, and opens defined-risk verticals as single multi-leg orders through Alpaca's official CLI — no SDK, so every broker call is a string you can replay in a terminal. It does not predict direction. It trades a structural claim: options are usually priced above the volatility that arrives. Per symbol it compares ATM implied vol against 20-day realized vol — sell a credit spread above 1.30, buy a debit spread below 0.85, stand aside in between. Standing aside is the common case by design. The hard part is that the denominator can break. Close-to-close realized vol cannot tell a jump from volatility, so one earnings gap inflates it, the ratio reads low, and options look cheap — while implied was already crushed by that same event. Both errors point the same way, so the agent is likeliest to buy premium exactly when it is least worth it. It did: NVDA at IV/RV 0.63 off a 45% realized reading. The window is now read twice, once with its largest day removed, and a signal must survive both. The check can only remove a trade, never create one. Five risk gates need a unanimous pass: drawdown 10%, portfolio 25%, per-underlying 5%, three trades a day, 2% per trade. Sizing uses the price the order is actually sent at, not the midpoint — the concession toward the market always moves toward more risk. One LLM sits at the end and can only veto a trade the rules already approved. It cannot propose, size, pick strikes, or overturn a gate, and it fails open. Everything here rests on the decision path staying reproducible from its inputs. The audit trail is the deliverable: 52 decisions logged, 41 of them refusals with stated reasons. A separate report joins every entry to its result and refuses to call seven trades evidence. Seven defects reached the live account, all documented with their numbers, including a failed cancel that tripled two positions. 309 tests, no network.
4 Sep 2026
Adaptive Routing Agent
Our AMD Developer Hackathon Track 1 entry is a general-purpose AI agent, packaged as a Docker image, that answers natural-language tasks across eight domains: factual knowledge, multi-step math, sentiment classification, summarization, named-entity recognition, code debugging, logical deduction, and code generation. Submissions must first pass an LLM-judge accuracy gate; those that pass are then ranked by the fewest total tokens used, so token efficiency, not raw accuracy, is what wins. Our design is a hybrid router. Each task is first classified almost for free with layered regular expressions, falling back to a tiny two-token model call only when the intent is genuinely ambiguous. The router then sends the task to the smallest Fireworks AI model that reliably clears the accuracy bar for that category: lightweight models handle sentiment, NER, and short summaries, while larger models are reserved for math, logic, and code. Model tiers are derived at runtime from the allowed-models list, so nothing is hardcoded. Prompts are deliberately terse, output length is capped per category, and every call runs concurrently through the judging proxy under a strict internal deadline, so valid results are always written on time.
13 Jul 2026
