.png&w=256&q=75)
🤝 Top Collaborators
🤓 Latest Submissions
EchoGrasp — The Robot That Knows When It Failed
EchoGrasp is a voice-directed bimanual robot-arm system built for the Intel Robotics track, stacked with the Speechmatics Bonus Award. Most robotics demos show a robot that works. EchoGrasp shows one that knows when it didn't, and fixes itself. You speak a command; Speechmatics transcribes it live; an LLM (via Groq) turns it into a strict, typed action plan; two simulated Franka Panda arms (PyBullet) execute it step by step; and after every placement, a real Intel OpenVINO INT8 model inspects a camera frame to judge whether the object actually landed where intended. When a step fails, the system says so in plain language ("gripper closed on nothing") and automatically re-plans with a corrected approach before retrying, up to a bounded limit. A live Honesty Dashboard shows the transcript, plan, per-step status, latency, and a running session success rate in real time. We're upfront that the OpenVINO model, trained on ~900 synthetic frames, measures roughly 60-70% accuracy alone, so the pass/fail call is confirmed against the simulator's own ground truth rather than the model's raw output. That model still runs genuine inference every placement step, and its confidence is what's reported live. In real warehouses, the expensive failures are the silent ones a system never notices; EchoGrasp is a small, working demonstration of catching those before they become someone else's problem.
16 Sep 2026
The Specialist-Options market-maker+convexity desk
Most trading agents guess direction — will a stock go up or down over a few days? Over a short competition window, that's close to a coin flip. The Specialist doesn't guess. It runs two independent strategies side by side on Alpaca paper trading. Specialist Mode acts like a real market maker: it posts resting limit quotes on liquid, near-the-money options across a small basket (SPY, QQQ, AAPL, NVDA, TSLA), capturing the bid-ask spread whichever way the underlying moves. Every fill triggers an immediate equity hedge, keeping the book close to delta-neutral. Convexity Mode runs in parallel, selling defined-risk vertical spreads and iron condors when implied volatility looks rich, so the account keeps generating real trading activity even on a quiet day for market-making fills. An LLM (via Featherless) decides which symbols to quote and how wide, reads scheduled macro events and live news sentiment into a risk posture — but it never places an order. A deterministic risk gate sits underneath every decision: hard caps on portfolio delta/vega/gamma, a daily drawdown circuit breaker, a kill switch, and reconciliation logic that specifically guards against Alpaca's documented ~10% random partial-fill rate on multi-leg orders, which can otherwise leave a spread as an unbounded naked position overnight. Everything is written to an append-only SQLite ledger and rendered on a static dashboard with zero credentials in the browser. Several of its safety rules exist specifically because running it live — not reading documentation — surfaced real platform constraints (Alpaca rejects naked short calls; it blocks a resting buy and sell on the same contract at once) that shaped the final design.
4 Sep 2026
