.png&w=256&q=75)
🤝 Top Collaborators
🤓 Latest Submissions
ReKit — Visual Recovery Intelligence
Incomplete returns can contain complementary parts, but a plausible visual match is not enough to rebuild a compatible kit. ReKit connects uncertain visual evidence to accountable recovery decisions. The prototype preserves original images and hashes under unique physical-component identities. Local Qwen3-VL-2B-Instruct inference on NVIDIA hardware records observations, uncertainty and source references. Humans verify or correct those claims; approved BOM and revision rules determine compatibility. CPU OR-Tools CP-SAT compares rebuilds, liquidation, donor retention and holds. Atomic reservations reject conflicting allocations, and explicit completion records preserve kit lineage. Two separate demonstrations show this boundary. Original Amazon Berkeley Objects catalog photographs support actual local visual inspection in simulated intake; they are not customer-return photographs. A scripted synthetic case uses three returns and four components: revision B is rejected, a new close-up requires review, and a recorded A correction enables two synthetic kit completions. Under baseline assumptions the modeled rebuild value is $175, compared with $102.50 whole-return liquidation and $16 noncash retention. Raising labor costs changes the preferred decision to $85 liquidation. These are scenario estimates, not sales or measured savings. ReKit is a validated local prototype, not a production deployment. Human semantic accuracy has not been measured. Actual AMD-hosted inference and application integration remain pending after two HTTP 400 attempts; NVIDIA execution does not satisfy AMD qualification. The contribution is the auditable connection between evidence, human authority, constrained allocation and controlled execution.
18 Oct 2026
VoxOps — Voice. Evidence. Safe Recovery.
VoxOps is an evidence-gated AI Voice Incident Commander built with the AssemblyAI Voice Agent API. Production incidents are stressful because engineers must quickly inspect service health, logs, deployments, dependencies, and possible recovery actions while avoiding decisions that could make the outage worse. VoxOps allows an engineer to investigate an incident using natural voice commands. AssemblyAI provides the live voice interaction, transcription, turn detection, agent reasoning, structured tool calling, and spoken responses. The important difference is that VoxOps does not blindly execute operational commands. Infrastructure facts are gathered through structured incident tools, including service metrics, logs, deployment history, and dependency health. The resulting evidence is used to form a diagnosis and validate any requested recovery action through deterministic safety policies. If a rollback is supported by the evidence, VoxOps creates a recovery proposal and sends it to a Human Safety Gate. The AI cannot bypass this gate, and the operator must explicitly approve or reject the action. If the requested action is not supported by the evidence, VoxOps blocks it before the Human Safety Gate even opens. For example, when the Payment API becomes unhealthy because of an external AtlasPay Gateway failure, a request to roll back Payment API to v3.7 is rejected with ROLLBACK_NOT_VALIDATED. The Payment API remains unchanged because the evidence shows that rollback would not address the actual root cause. VoxOps also provides recovery verification, an incident timeline, automated postmortem reporting, deterministic safety evaluation, and a live voice-agent benchmark. The AssemblyAI voice and tool-calling path is live. The production services, infrastructure telemetry, and recovery execution in this hackathon prototype are simulated so the safety architecture can be demonstrated without modifying real production infrastructure.
30 Sep 2026
ParityLens: Preprocessing Parity Debugger
Computer-vision pipelines often fail in a subtle way: the final tensor can have the correct shape and dtype, while its semantic content is wrong. A channel-order change, missing scaling step, or incorrect normalization can silently alter what the model receives, and simple structural checks may still pass. ParityLens is a developer tool for finding exactly where that preprocessing drift begins. It runs a trusted reference pipeline and a candidate serving pipeline side by side, records their outputs at five boundaries: decode, geometry, scaling, normalization, and model input, and stops at the first comparable stage where parity fails. The interface then shows the actual values, shapes, dtypes, maximum difference, source location, and defect classification. Our main demo preserves a real historical RGB/BGR bug. The reference and candidate both pass shape and dtype checks, but ParityLens detects the first semantic mismatch at decode. The recorded probe changes from [180, 120, 20] to [20, 120, 180], exposing the red/blue channel swap. IBM Bob then diagnosed the issue and produced the minimal BGR-to-RGB repair. The repaired candidate is verified again with the same frozen contract, comparator, verifier, and fixtures. ParityLens is designed for computer-vision and ML engineers moving code from experiments to serving, refactoring preprocessing code, migrating frameworks, or comparing edge and cloud pipelines. Users open the live Streamlit app, reproduce the historical bug, inspect the first failing boundary, verify the Bob repair, and explore controlled scaling and normalization failure cases. What makes ParityLens different is the combination of first-boundary localization, numerical evidence, preserved before/after provenance, and verification under unchanged checks. In our deterministic synthetic evaluation, ParityLens detected 45/45 supported defects, localized all 45 first failing boundaries, and produced 0/15 false positives on clean controls. A final shape/dtype-only baseline detected 0/45 of those defective cases. The current prototype is intentionally narrow and honest: it uses synthetic, contract-constrained inputs and does not claim production-wide coverage. Its purpose is to make hidden preprocessing mismatches inspectable, reproducible, and easier to repair.
27 Sep 2026
TableGuard-Lite: Safe Bimanual VLA Control
TableGuard-Lite is a safety-aware closed-loop Vision-Language-Action (VLA) system for bimanual robotic table setting. It combines natural-language instructions, three RGB camera views, and a 12-joint robot state to generate coordinated actions for two SO-101 robot arms using SmolVLA. Instead of executing long open-loop trajectories, the system predicts short action chunks, executes only a small portion, and then replans from fresh observations. A runtime safety supervisor checks joint-rate limits, unexpected contacts, collisions, and scene preservation, stopping unsafe behavior before it can propagate. TableGuard-Lite also introduces recovery-oriented learning: states encountered during failed closed-loop execution are captured and reused for targeted fine-tuning. The project includes a successful full reference manipulation sequence, learned closed-loop control experiments, failure-driven recovery training, three-camera perception, and safety-supervised execution. OpenVINO export, numerical-parity verification, and Intel CPU inference were also successfully demonstrated. TableGuard-Lite shows how Physical AI systems can combine VLA intelligence with safety, recovery, and deployment verification for more reliable real-world robotic manipulation.
16 Sep 2026
