Use cases · AI Annotation Lab

30 concrete annotation workflows XUmmia ships.

Every use case below runs on the same calibrated pipeline - versioned guidelines, golden-set benchmarking and blind consensus review. Pick the workflow that matches your data and we'll return a scoping doc in 72 hours.

Image annotation

Vision datasets, pixel-honest.

From single-label classification to dense segmentation on culturally-grounded imagery.

Image classification

Image classification

Multi-class labels for retail SKUs, agricultural crops or medical scans that off-the-shelf datasets can't cover.

XUmmia delivers
Balanced class distributions with edge-case sampling and a per-class confusion audit.

2D bounding boxes

2D bounding boxes

Object detection for street scenes, satellite imagery and warehouse logistics with tight IoU targets.

XUmmia delivers
≥0.94 IoU on golden sets, double-review on occluded objects, JSONL + COCO delivery.

3D cuboid & LiDAR boxes

3D cuboid & LiDAR boxes

Autonomous-driving perception stacks need 3D extents, orientation and velocity flags - not just 2D.

XUmmia delivers
Cuboid annotation on point clouds with per-frame temporal consistency checks.

Semantic segmentation

Semantic segmentation

Pixel-level masks for medical imaging, land-use satellite maps and content-moderation pipelines.

XUmmia delivers
Polygon-first workflow with reviewer-side mask diffing and boundary error heatmaps.

Instance segmentation

Instance segmentation

Counting and separating overlapping instances - cells under a microscope, produce on a market table.

XUmmia delivers
Per-instance mask + class labels with instance-count reconciliation across reviewers.

Keypoint & pose estimation

Keypoint & pose estimation

Athlete tracking, physiotherapy motion capture and gesture recognition need anatomically stable keypoints.

XUmmia delivers
23-point skeletons with visibility flags and inter-keypoint distance sanity checks.

Text & NLP annotation

Language models learn from labelled language.

Structured NLP labelling across 40+ languages and domain rubrics your model actually needs.

Named entity recognition (NER)

Named entity recognition (NER)

Domain entities - Nigerian legal citations, Swahili drug names, financial tickers - that generic NER misses.

XUmmia delivers
Span-level labels with nested-entity support and per-entity precision/recall dashboards.

Text classification & intent

Text classification & intent

Customer conversations, support tickets and social posts labelled for intent, urgency and product line.

XUmmia delivers
Multi-label taxonomies with confidence bands and reviewer disagreement analytics.

Sentiment & emotion labelling

Sentiment & emotion labelling

Beyond positive/negative - sarcasm, code-switching and dialect-specific affect that Western models miss.

XUmmia delivers
6-axis emotion rubric calibrated per language with reviewer-panel adjudication.

Summarization gold sets

Summarization gold sets

Reference summaries for evaluating LLM abstractive quality across long-form documents.

XUmmia delivers
Multi-reference summaries with faithfulness, coverage and conciseness scores per pair.

Translation quality (MQM)

Translation quality (MQM)

African-language translations rated for accuracy, fluency, terminology and register at MQM granularity.

XUmmia delivers
Sentence-level MQM error typing with severity weights and bilingual reviewer chains.

Audio annotation

Speech and sound, ground-truthed.

ASR training data, speaker analytics and audio-event labelling on the languages your users speak.

Speech transcription (ASR)

Speech transcription (ASR)

Verbatim transcripts for low-resource languages and code-switched conversations with disfluency preservation.

XUmmia delivers
Timestamped transcripts with WER benchmarking against multiple reviewers.

Speaker diarization

Speaker diarization

Multi-speaker call recordings and podcast episodes labelled with speaker turns and identity tags.

XUmmia delivers
Segment-level speaker labels with turn-boundary consensus across 3 reviewers.

Audio-event tagging

Audio-event tagging

Environmental sound classification for wildlife monitoring, urban safety and industrial acoustics.

XUmmia delivers
Weakly-supervised event tags plus strong onset/offset labels on a curated subset.

TTS voice rating

TTS voice rating

Synthetic voices need naturalness, intelligibility and pronunciation-error ratings by native listeners.

XUmmia delivers
5-point MOS with per-phoneme error notes and native-speaker reviewer panels.

Video annotation

Temporal labels, frame-accurate.

Tracking, event detection and moderation on real-world video from the markets, roads and clinics of the continent.

Multi-object tracking (MOT)

Multi-object tracking (MOT)

Persistent identities across occlusion, entry and exit for surveillance, sports and retail analytics.

XUmmia delivers
Tracklet IDs with HOTA scoring and per-track review of ID switches.

Action & event segmentation

Action & event segmentation

Temporal segments for actions in sports, procedural videos, safety incidents and clinical exams.

XUmmia delivers
Frame-accurate segment boundaries with a start/end tolerance rubric per event class.

Content-moderation labels

Content-moderation labels

Policy labelling of user-generated video - safety, minors, IP and regional-law categories.

XUmmia delivers
Multi-policy taxonomy with reviewer well-being safeguards and daily rotation caps.

Rating & preference

The judgment signal behind RLHF.

The pairwise, rubric-driven ratings that turn raw responses into an alignment signal.

Pairwise preference (RLHF)

Pairwise preference (RLHF)

Two LLM responses, one prompt - decide which is better and why, at scale, without collapsing into noise.

XUmmia delivers
A > B decisions with rubric-tagged rationales and per-rater calibration to a golden set.

Rubric scoring for LLM eval

Rubric scoring for LLM eval

Score generations on factuality, safety, helpfulness and local relevance with reproducible criteria.

XUmmia delivers
Per-axis Likert scores with reviewer-panel adjudication and score-drift monitoring.

Domain expert annotation

Vetted MDs, JDs, CPAs, PhDs and senior engineers.

High-stakes labelling and eval by credentialed, NDA-bound experts - the same tier of talent Turing and Outlier lead with.

Medical & clinical

Medical & clinical

Clinical-note QA, diagnosis reasoning and medical coding need reviewers who can spot a wrong differential, not just a wrong token.

XUmmia delivers
MDs and RNs reviewing generations with SNOMED / ICD-10 coding and safety-flag rubrics.

Radiology & medical imaging

Radiology & medical imaging

X-ray, CT and MRI finding-level labels and segmentations require radiologist-grade judgment on subtle abnormalities.

XUmmia delivers
Board-eligible radiologists producing finding-level labels and pixel-accurate lesion masks.

Legal & contract review

Legal & contract review

Clause extraction, case-law reasoning and jurisdictional QA fail without a JD who reads the way courts do.

XUmmia delivers
JDs and licensed attorneys extracting clauses, tagging obligations and adjudicating case-law reasoning.

Finance & accounting

Finance & accounting

Financial-doc extraction, tax and audit reasoning and XBRL-style tagging demand CPA and CFA-level accuracy.

XUmmia delivers
CPAs and CFAs labelling 10-Ks, transaction data and tax positions with audit-trail rationales.

Scientific & STEM reasoning

Scientific & STEM reasoning

Math proofs, physics and chemistry problem QA and paper-grade reasoning need PhD-level reviewers.

XUmmia delivers
PhDs adjudicating step-level math, physics and chemistry reasoning with error taxonomies.

Senior code review

Senior code review

Architecture critique, security review and PR-grade rationales fall apart with junior labellers.

XUmmia delivers
Senior engineers writing PR-grade critiques, security notes and refactor rationales.

AI Governance, Risk & Compliance

AI Governance, Risk & Compliance

Model-risk assessment, NIST AI RMF / ISO 42001 mapping and EU AI Act classification need GRC specialists, not generalist raters.

XUmmia delivers
AI GRC leads and auditors producing risk registers, bias audits and policy red-team reports.

Quality control

The layer above every label.

XUmmia's own QA is a product - golden sets, reviewer calibration and public agreement metrics on every project.

Golden-set benchmarking

Golden-set benchmarking

Reviewers drift, rubrics change, model outputs shift - teams need a stable ground truth to measure against.

XUmmia delivers
Rolling golden sets with per-reviewer accuracy scoring and auto-retirement of stale items.

Inter-annotator agreement (κ / α)

Inter-annotator agreement (κ / α)

Sign off on a rubric only after you know reviewers agree - Cohen's κ, Fleiss' κ, Krippendorff's α.

XUmmia delivers
Live agreement dashboards with per-class breakdowns and rubric revision loops.

Reviewer calibration & onboarding

Reviewer calibration & onboarding

New reviewers drift from the rubric until they're calibrated - and drift again when guidelines change mid-project.

XUmmia delivers
Structured onboarding on gold items, per-reviewer accuracy tracking and re-calibration gates on every rubric revision.

Your workflow isn't on this page?

We've shipped labelling pipelines that never made this list. Send us the brief - you'll have a scoping doc, cost model and expert roster in 72 hours.