🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
Voice Action Gate
Voice Action Gate is a security layer for AI voice agents that prevents sensitive actions from being executed based on guessed, missing, or ambiguous intent. It verifies that required parameters were explicitly provided, checks risk and confirmation requirements, and blocks unsafe requests before they reach protected tools. The system separates AI understanding from authorization, maintains an audit trail, and demonstrates secure workflows such as deployments, database actions, and simulated fund transfers. Its goal is simple: AI may understand what you mean, but it should only act on what you clearly authorized. :chatgpt-content-reference{index="0"}
30 Sep 2026
TaskForge VLA
TaskForge VLA is a simulation-first embodied AI system built for the AI Infra Summit Hackathon 2026, targeting the Intel Bimanual VLA Manipulation with Multi-Modal Reasoning track. Users can give natural-language commands such as “Pick both blocks” or “Put the red block near the blue one.” TaskForge interprets the instruction with an OpenRouter-powered LLM, converts it into a typed, structured goal, and falls back to a deterministic parser if needed. The system observes the simulated world, generates a symbolic plan, and uses a reachability-aware bimanual coordinator to decide which arm should perform each action. Before execution, a deterministic validator checks objects, actors, dependencies, grasp states, and supported actions. Validated plans are executed in MuJoCo using deterministic primitives such as pick, place, and move-home, with analytical inverse kinematics controlling the arms. After execution, TaskForge re-observes the world and verifies whether the original goal is satisfied using predicates such as `left_of`, `right_of`, `near`, and `grasped_by`. A key feature is closed-loop recovery. If an action fails, TaskForge preserves completed work, observes the updated world, creates a residual goal for only the unfinished task, replans, validates the new plan, and continues execution. The project also includes a Streamlit dashboard for task input, plans, world state, execution status, recovery events, and final verification, plus a live MuJoCo viewer for visual robot simulation. Core principle: AI interprets. Structured systems validate. Deterministic robotics executes. The environment decides success.
16 Sep 2026
.png&w=828&q=75)
Hedgify
Hedgify is an autonomous, multi-agent portfolio insurance system built for the Alpaca AI Trading Agents Hackathon (lablab.ai). It solves a problem every equity investor knows intimately: the panic-sell. When markets crash, retail investors liquidate their holdings to stop the bleeding — crystallizing losses, triggering tax events, and inevitably missing the rebound. Hedgify replaces that reflex with insurance: instead of selling your stocks when danger approaches, the system automatically buys protective put options — capping your downside while you keep every single share and ride the recovery. The system is architected as a Supervisor–Worker multi-agent pipeline. The Monitor Agent polls live portfolio data from Alpaca's paper trading API, tracks the all-time peak value, and computes real-time drawdown the moment any poll lands. When the drawdown crosses its threshold, the Monitor persists a snapshot, fires a structured alert into the database, broadcasts it over WebSocket to any connected dashboards, and hands off to the Hedge Executor Agent. The Executor then runs its gauntlet: an idempotency guard first checks whether the symbol already carries active protection (sixteen duplicate alerts become exactly zero duplicate orders); the live kill-switch is consulted, letting an operator disable autonomous trading mid-crisis without restarting anything; then the agent queries Alpaca's live options chain over its dedicated data API, matches the target strike (5% out-of-the-money by default) to the closest actually listed contract, validates expiry against the settings table, checks the premium against the configured budget, and places a real order through Alpaca's trading API — then polls asynchronously for the actual fill price, recording genuine market premiums rather than estimates.
4 Sep 2026
LeakLogic AI
LeakLogic AI is an explainable AI business-auditing platform built for small and medium-sized businesses that need to understand where profit is being lost. Users upload sales, refunds, supplier-cost, and inventory CSV files through a modern web dashboard. The FastAPI backend validates and normalizes the data, then runs four deterministic Python and pandas detectors for refund anomalies, discount leakage, supplier margin compression, and inventory drag. Each finding includes the affected entity, measured change, supporting evidence, confidence score, estimated dollar impact, likely cause, and recommended action. Findings are ranked by financial impact so decision-makers can focus on the most important issues first. After the calculations are completed, Fireworks AI transforms the verified structured findings into a concise executive narrative. The model is instructed not to invent figures, causes, evidence, or recommendations. If the AI service is unavailable, LeakLogic AI automatically returns a deterministic fallback summary, ensuring the audit still completes. The application is fully containerized with Docker and Docker Compose. The production frontend is deployed on Netlify, while the FastAPI backend is deployed on Vercel. The project uses Fireworks AI API access from the AMD hackathon ecosystem and was developed for the Unicorn Track as a product-focused AI solution. LeakLogic AI turns disconnected operational data into a clear answer to one business question: where are we quietly losing money, why does it matter, and what should we review next?
13 Jul 2026
