🤓 Latest Submissions
Bollard: the gate between hearing and hands
Bollard is the gate between hearing and hands: a voice-to-action interlock for robot arms. Every spoken command is transcribed by Speechmatics Realtime, parsed by a deterministic grammar, and judged by a safety policy before anything moves. Allowed commands execute on a bimanual dinner-table simulation; denied commands return named codes and human-readable reasons (the saucepan is hot, the candle is lit, glass never gets thrown). Every command, allowed or denied, appends to a SHA-256 hash-chained receipt recording what was heard, the transcription confidence, the parsed intent, the verdict with reasons, and the scene fingerprint afterwards. One click re-hashes the entire chain and prints VALID or the exact serial where it breaks. Why: the field ships hearing (speech-to-text) and hands (VLA policies and arms), but nothing ships the gate between them. Enterprise robot deployment stops at exactly this problem: certification and safety review demand an auditable artifact of what the machine acted on. Bollard is that artifact, and it is deliberately judgeable: open the cockpit, press a command chip, watch HOLD turn into ALLOW while the arm moves, then try to break it with a hazard command, then hit Verify chain. Engineering: Next.js + TypeScript on Vercel (static + one serverless function), Speechmatics Realtime over WebSocket with 120-second JWTs minted server-side so the API key never reaches the browser, custom vocabulary boosted for table terms, a deterministic intent grammar and policy engine with no LLM in the command path, and a canvas simulation with two arms, a bollard and a rope that holds while the policy reads, pays out on allow, and snubs on deny. Tracks: Intel online (Bimanual VLA Manipulation with Multi-Modal Reasoning, dinner-table challenge option, simulation-first) plus the Speechmatics Best-Use bonus award. Solo builder, agents-augmented.
16 Sep 2026
Tape: the proof-carrying options desk
Tape is an autonomous options desk on Alpaca paper that prints a serial-numbered receipt for everything it does, then reconciles its own journal against the broker every cycle. Live: tape.up.railway.app AI logic: a Groq analyst reads desk state and the option chain and proposes defined-risk credit spreads with its thesis attached. The model proposes only: sizing, gates and flattens are deterministic code. A nightly scribe grades closed trades into lessons that feed later cycles (10 on record, including the ones that hurt). Risk gates: 13 deterministic gates run before any order exists: 2.5% NAV per trade, 20% aggregate open risk, 4% daily halt, delta band 0.10-0.40, max 5 DTE, liquidity, pin guards, Friday flatten. 33 gate runs on record: 11 approved, 22 blocked with the exact gate that killed each. Alpaca infrastructure: the agent speaks to Alpaca through the official MCP server over stdio; every tool call is journaled (15,138 calls with full payloads). Orders execute as multi-leg submissions with fill verification before journaling. A reconcile pass every cycle compares the journal against get_all_positions: clean stamps on record for every cycle of the submission account (PA37YPE0PC51). The spine is honesty: on day two the desk froze, held 0-DTE spreads into expiry and the first account died at -$54,152. It was archived as a receipt chain, the failure became lessons, and the fresh book now trades with fill verification and a public reconcile wall. A Sep 3 expiry left -200 SPY short stock on the book at submission time: receipted, marked live, flattened by the Friday rule.
4 Sep 2026
