Evaluation interfaces for LLMs and AI agents

Layout, OCR, tables, and classification usually live in four different tools, and field extraction happens somewhere else again. This document AI interface puts them in one task.

Annotators draw bounding boxes around document elements, select text spans and words backed by hOCR positions, mark up table rows and columns with per-cell labels and merged cells, classify the document and each page, and link related regions, such as a caption to its figure. An extraction panel binds labeled regions to named fields so you can preview the structured record before saving.

Works on PDFs and single-page images. The region labels, document types, and page types are parameters you replace with your own.

Multi-turn failures accumulate. A turn that looks fine on its own is wrong given what the user said three turns earlier, so the reviewer needs the whole transcript in view while grading one turn. This interface does that.

Each assistant turn gets four 1 to 5 rubric scores, issue flags, notes, and text spans marked as evidence, claim, or correction; attached assets such as images, code, tables, and audio render inline. A conversation-level rubric at the end computes a verdict (Excellent, Good, Mixed, Poor) and takes summary notes.

An agent run is not one answer to grade. It is reasoning, tool calls, subagents, artifacts, and errors, and the failure is usually somewhere in the middle. This interface lays the whole trace out so a reviewer can find it.

Steps render in order with their arguments, results, tokens, and latency, and a minimap sized by latency, tokens, or cost shows where the run spent its budget. Reviewers rate each step, set a verdict, score a configurable rubric (helpfulness, faithfulness, efficiency, tool use, and instruction following by default), tag failure modes like hallucination or wrong tool, and write a critique.

Each section saves independently, so partial reviews are fine. Verdicts, step ratings, failure modes, and rubric items are all parameters.

A policy rollout is judged on video, but the useful output is a timeline: where each subtask started, where the mistake happened, what the coaching note should say. This interface keeps both in view.

Annotators watch top, wrist, and other camera streams in a single, gallery, or main-plus-side layout, scrub with frame stepping and speed control, and edit segments from the timeline toolbar: add a subtask, split at the playhead, mark a mistake, pin a subgoal. A metadata panel holds the 1 to 5 quality rating, control mode, speed bin, and robot ID.

A DICOM file already carries the structure — the trick is getting it on screen in a form someone can actually work with. This interface reconstructs the study into frontal and side views you can scroll slice by slice, with presets that switch between the file's own header defaults, the full unfiltered pixel range, and standard grayscale ranges for bone, soft tissue, or teeth. Labels are read directly out of the DICOM, so every encoded structure arrives already segmented, ready to refine with brush and erase tools or dial back with an overlay-strength control. A WebGL 3D view renders the whole scan, and selecting a structure there carries that selection back into the slice view, with the file's full header data alongside for reference.

Most perception teams annotate the same scene four times: once for cuboids, again for camera views, again for tracking, and once more for review. This interface keeps all of it in one workspace.

Annotators work the 3D point cloud alongside a top-down view and a camera panel that follows whatever cuboid is selected. Keyframe interpolation carries objects across frames, so adjusting one position updates the frames around it instead of redrawing each one — which is where identity swaps usually creep in. Model predictions load as editable suggestions to accept, adjust, or reject. Validation runs inside the task, so blocking errors surface before completion rather than in a separate audit pass weeks later.

Handles COPC files in the tens of millions of points, with cuboids, points, lines, lanes, and segmentation masks side by side.

A wearable session is five signals that have to be read against each other: heart rate climbing after the accelerometer, HRV dropping, respiration following, SpO2 holding until it dips. This interface puts all five on one time base with a context track underneath, so precomputed activity, sleep, and rhythm segments stay visible while you work, and outdoor sessions carry a GPS route beside the signals.

Labeling splits across three tabs, and the vocabularies ship as defaults you replace: Running through Sedentary for activity, Awake through Deep Sleep for sleep, Sinus Rhythm through AFib for rhythm, each label with a color and a hotkey.

The task fields are parameters too, so the interface reads whatever your sensor, segment, and route columns are called instead of forcing your data into its names.

Your documents do not look like Docling's training data: a lab report or an internal invoice has little in common with the public PDFs a converter learned on. This interface gives you a place to see where that breaks.

You draw bounding boxes around document elements, layer reading-order paths across them, and correct merged regions and table structure, with the DocLang XML updating in a sidebar as you work. Predictions from the Docling ML backend load as a starting point, and the result exports as a .dclx archive carrying both the structure and the page images.