▸case-22 We need an automated vision pipeline that inspects utility pole equipment photos, locates transformers, and directly extracts structured serial numbers and maintenance status via visual reasoning. Write an architecture specification for this visual question answering system. Format as a technical design document. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-17 We are building an ergonomic analysis tool that needs to extract full 3D human body and object meshes from standard single-view RGB video feeds taken in workplace environments. Design a foundation-model-driven pipeline for 3D spatial reconstruction from monocular inputs. Format as an architecture document. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-05 We are designing a weatherproof IP67 enclosure and mechanical mount in SolidWorks for a dual-camera stereo rig mounted on high-vibration agricultural equipment. Please draft a mechanical design document detailing material selection, thermal dissipation strategies for passive cooling, and mechanical isolation to handle shock loads. Format as a mechanical engineering design report with numbered specifications. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-02 We need to automate visual inspection in a manufacturing warehouse where workers identify damaged items using spoken or written descriptions (like 'the scratched silver cylinder on the left side'). Please design an end-to-end vision pipeline that leverages modern foundation models to segment objects directly from text queries. Provide a structured design report containing a pipeline overview diagram, a complete Python implementation sketch demonstrating model initialization and inference, and a recommendations table covering common operational edge cases. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-13 In a high-speed bottling plant, rapid conveyor movement causes significant motion blur in captured frames, degrading segmentation mask consistency across sequential frames. Write an operational mitigation guide addressing camera hardware settings and model temporal features to maintain tracking quality. Format as an operational guide. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-07 We are building an interactive video annotation system for satellite imagery where users enter category names and the system must locate and create pixel-precise boundaries for thousands of structures per frame. Using an off-the-shelf single heavy foundation model directly for every candidate location creates severe compute bottlenecks. Write an architectural proposal describing a two-stage hybrid pipeline strategy. Format as an architecture document with component breakdown and dataflow steps. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-11 We are preparing a real-time object detection model deployment pipeline targeting TensorRT and ONNX Runtime on low-power NPU hardware. Engineers often encounter performance bottlenecks when exporting traditional detectors due to complex custom C++ or CUDA NMS plugins. Draft an export optimization protocol that eliminates these post-processing dependencies. Format as a step-by-step export protocol. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-12 We need to convert 2D video frames captured from a handheld inspection camera into a 2.5D or 3D spatial representation to measure object heights and distances. Draft a technical workflow proposal describing how to combine monocular depth prediction with geometric camera math. Format as a software pipeline specification. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-20 An ML deployment engineer is updating an edge export script for an updated real-time object detector targeting low-power microcontrollers. They are considering keeping their legacy Distribution Focal Loss (DFL) post-processing layer in the exported ONNX model. Draft an architectural review recommending whether to retain or strip legacy DFL structures during modern edge model deployment. Format as a code review advisory. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-21 A drone surveillance system requires continuous real-time detection, tracking, and pixel-level mask generation of moving vehicles across dynamic urban environments. Using separate models for detection, segmentation, and tracking causes frame drops. Design a unified model architecture solution. Format as a system architecture blueprint. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-04 We are developing a custom camera driver module in C for an embedded Linux platform communicating with an Omnivision sensor via V4L2 and I2C registers. Please write a C driver snippet and step-by-step register configuration guide to set exposure time and gain controls. Structure the output as an embedded C development guide with code snippets and kernel IOCTL explanations. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-18 Our machine learning team is fine-tuning custom detection models for IoT security cameras. Using standard AdamW SGD optimization on our large custom dataset requires excessive epoch counts before convergence. Propose a modern optimizer choice and training strategy tuned for quick convergence on custom edge vision architectures. Format as an ML training optimization guide. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-19 An electronics remanufacturer handles hundreds of distinct circuit board variants without labeled training sets for each. They need to locate and isolate specific components during inspection using natural language descriptions without training custom detectors for every new part number. Write an operational pipeline proposal detailing foundation model selection. Format as a technical solution brief. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-08 We are training an automated optical inspection (AOI) model to spot micro-cracks and missing 0201 surface-mount resistors on printed circuit boards. Standard object detection losses often fail on such small features in high-resolution images. Draft a training optimization plan detailing specific loss functions, assignment strategies, and optimizers for fast convergence on small objects. Provide structured bullet points and PyTorch configuration snippets. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-14 We are setting up a high-precision multi-camera rig for automated dimensional metrology in automotive manufacturing. Standard manual calibration grids often suffer from corner distortion and sub-pixel inaccuracy. Design a sub-pixel camera calibration protocol detailing target pattern selection and software algorithms. Format as an engineering procedure document. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-15 An autonomous indoor delivery robot needs to build a map of unknown hospital corridors and track its real-time position without relying on external GPS or beacon infrastructure. Outline a visual localization and mapping framework suitable for this mobile platform. Format as a spatial perception design spec. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-03 Our team is designing a spatial perception stack for a small indoor mobile robot that needs depth estimation and 3D scene understanding using a single RGB camera setup. Please write a comprehensive architecture document outlining how to integrate monocular depth estimation with geometric calibration and spatial mapping. The output should be formatted as a software design document with an architecture breakdown, a step-by-step data processing workflow, and an operational advice table for handling environmental challenges like dynamic motion and changing lighting. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-01 I am building a real-time defect detection system for high-speed industrial conveyor lines running on embedded low-power edge hardware. Please draft a detailed technical specification and deployment guide for this computer vision pipeline. The document should include an architectural diagram overview, recommendations for real-time model selection and post-processing strategies to keep latency minimal, and an export configuration walkthrough for hardware accelerators. Format your response as a formal engineering spec with section headers, structured bullet points, and code snippets where appropriate. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-06 We need to train an acoustic speech recognition system fine-tuned for high-noise factory environments to capture verbal worker commands. Please draft a machine learning workflow guide for fine-tuning acoustic speech models using audio augmentation and spectrogram features. Format as an ML pipeline architecture proposal. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-09 Our team wants to run promptable text-to-mask segmentation on an edge GPU workstation with limited 8GB VRAM, but full-scale SAM 3 models hit out-of-memory errors during batch inference. Write a memory optimization guide outlining concrete techniques to reduce GPU VRAM footprint without retraining the model base. Format as a troubleshooting guide with bulleted recommendations. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-10 We are developing an automated document understanding system that needs to parse complex industrial schematics and answer natural language questions about specific component connections directly from visual inputs. Design a multimodal vision architecture that performs visual grounding and visual question answering (VQA) on high-resolution diagrams. Format as a technical design document. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-16 When operators query our text-guided segmentation pipeline with generic prompts like 'bolt' or 'wire', the model frequently selects adjacent unrelated hardware components. Write a prompt engineering guide for vision-language segmentation models to reduce visual false positives. Format as best-practices documentation with bad/good prompt examples. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |