Manuel Ogiez

Key Skills

Software

Clickworker

Data Annotation Tech

Labelbox

Lionbridge

Scale AI

Telus

Internal/Proprietary Tooling

Top Subject Matter

No subject matter listed

Top Data Types

Geospatial Tiled Imagery

Image

Text

Top Task Types

Audio Recording

Computer Programming Coding

Evaluation Rating

Mapping

Prompt Response Writing SFT

Freelancer Overview

Multilingual AI Trainer and Data Annotator with hands-on experience since January 2025 delivering high-quality training data for Large Language Models and speech systems in both English and French. I specialize in designing rigorous evaluation rubrics, ranking and red-teaming model outputs to surface hallucinations, policy violations, bias and safety risks. My work spans multimodal labeling—text, code, image, audio and geospatial POI—where I consistently exceed 98 % QA acceptance while meeting aggressive throughput targets. Notable contributions include crafting “LLM Judge” prompt templates that accelerated evaluation workflows, producing studio-grade voice recordings for TTS datasets, and mentoring distributed annotation teams through clear guidelines and calibration sessions. Combining technical acumen (Python basics, Git, modern annotation platforms) with a decade of caregiving and education experience, I bring a rare blend of empathy, pedagogical clarity and relentless attention to detail that elevates both data quality and team performance.

Labeling Experience

Bounding-Box Label Verification & Missing-Domain QA

TelusImageBounding BoxClassification

Task presents an image with pre-drawn bounding boxes and labels. Task 1 – Label Verification: select Correct or Incorrect for each label while the linked box is highlighted yellow. Task 2 – Missing Domain Verification: after label checks, decide if any of the predefined Domains (Document, Art, Landmark, Animals, Plants, Food, Sign, Storefront, Packaged Products, etc.) are present but un-boxed; if so, tick “Yes” and choose all missing Domains. Key rules: • Ignore people and body parts • Treat “Document” as a VIP domain with many sub-labels (menu, receipt, ticket, etc.) • Multiple unlabeled objects from the same Domain still count as “missing” • Use zoom and label-toggle tools to inspect small or occluded items.

Task presents an image with pre-drawn bounding boxes and labels. Task 1 – Label Verification: select Correct or Incorrect for each label while the linked box is highlighted yellow. Task 2 – Missing Domain Verification: after label checks, decide if any of the predefined Domains (Document, Art, Landmark, Animals, Plants, Food, Sign, Storefront, Packaged Products, etc.) are present but un-boxed; if so, tick “Yes” and choose all missing Domains. Key rules: • Ignore people and body parts • Treat “Document” as a VIP domain with many sub-labels (menu, receipt, ticket, etc.) • Multiple unlabeled objects from the same Domain still count as “missing” • Use zoom and label-toggle tools to inspect small or occluded items.

2025

Maps Search Evaluation – Search Relevance & Search 2.0 Rater

TelusGeospatial Tiled ImageryGeocodingMapping

Apply the Maps Search Evaluation Guidelines (June 2024) to judge each task: • Answer the query-level navigational question, then rate every result for Relevance (Navigational → Bad) • On Search 2.0 tasks, audit Name, Address & Pin accuracy after relevance scoring • Use distance ruler, viewport tools and comment boxes; demote for User-Intent or Distance/Prominence issues and leave concise English comments with sources when ratings fall below Excellent • Follow release-survey rules for adult content or technical blocks, otherwise complete and submit each survey

Apply the Maps Search Evaluation Guidelines (June 2024) to judge each task: • Answer the query-level navigational question, then rate every result for Relevance (Navigational → Bad) • On Search 2.0 tasks, audit Name, Address & Pin accuracy after relevance scoring • Use distance ruler, viewport tools and comment boxes; demote for User-Intent or Distance/Prominence issues and leave concise English comments with sources when ratings fall below Excellent • Follow release-survey rules for adult content or technical blocks, otherwise complete and submit each survey

2025

“xylophone_act” Dual-Role Voice-Data Collection Project

Scale AIAudioAudio Recording

Ongoing voice-data project that captures two complementary audio segments—Side A prompts and Side B responses—to improve speech and conversation models. Key activities: • Record high-quality, emotion-rich audio in varied home setups while preserving natural breaths. • Follow detailed Audio Setup Guide (microphone distance 6-12 in, pop filter, –12 dB to –3 dB peaks). • Adhere to the Delivery Guide to avoid robotic or scripted tone. • Perform 5-second silences between lines and export WAV files to project specs. • Review and fix transcripts so they match audio exactly, including interjections, pauses “...”, and non-verbal cues <[laughter]>.

Ongoing voice-data project that captures two complementary audio segments—Side A prompts and Side B responses—to improve speech and conversation models. Key activities: • Record high-quality, emotion-rich audio in varied home setups while preserving natural breaths. • Follow detailed Audio Setup Guide (microphone distance 6-12 in, pop filter, –12 dB to –3 dB peaks). • Adhere to the Delivery Guide to avoid robotic or scripted tone. • Perform 5-second silences between lines and export WAV files to project specs. • Review and fix transcripts so they match audio exactly, including interjections, pauses “...”, and non-verbal cues <[laughter]>.

2025

Chat-and-Compare Instruction-Following Evaluation (French)

Internal Proprietary ToolingTextEvaluation Rating

Continuous task focused on French-language prompt creation and response comparison: • Write a French prompt, receive two model responses, then rate them side-by-side. • Assess instruction-following, helpfulness and safety—factuality is no longer scored. • Provide a 3–5-sentence comparative insight explaining the ratings. • Edit the better response only when both are poor and the fix does not add new information or correct factual errors. • Ensure each conversation has ≥ 2 turns, avoid niche topics, respect an Oct 2023 knowledge cut-off, and vary prompt structures to boost diversity. • Flag factual mistakes via checkbox and describe them in the insights section when required.

Continuous task focused on French-language prompt creation and response comparison: • Write a French prompt, receive two model responses, then rate them side-by-side. • Assess instruction-following, helpfulness and safety—factuality is no longer scored. • Provide a 3–5-sentence comparative insight explaining the ratings. • Edit the better response only when both are poor and the fix does not add new information or correct factual errors. • Ensure each conversation has ≥ 2 turns, avoid niche topics, respect an Oct 2023 knowledge cut-off, and vary prompt structures to boost diversity. • Flag factual mistakes via checkbox and describe them in the insights section when required.

2025

French-Language LLM Judge-Instruction Design

Internal Proprietary ToolingTextEvaluation Rating

Created reusable instructions (“judges”) that evaluate LLM outputs while meeting strict guidelines: • Wrote all instructions in French. • Defined each judge’s objective, evaluation criteria, decision-making guidelines, explicit output format, and examples (≈50 % of cases). • Followed project rules — no niche or political topics, no external tools, knowledge cutoff ≤ Oct 2023, and mandatory structured outputs. • Ensured every judge could be applied to multiple user inputs and parsed automatically.

Created reusable instructions (“judges”) that evaluate LLM outputs while meeting strict guidelines: • Wrote all instructions in French. • Defined each judge’s objective, evaluation criteria, decision-making guidelines, explicit output format, and examples (≈50 % of cases). • Followed project rules — no niche or political topics, no external tools, knowledge cutoff ≤ Oct 2023, and mandatory structured outputs. • Ensured every judge could be applied to multiple user inputs and parsed automatically.

2025

Show more

Education

Work History