For employers

Hire this AI Trainer

Sign in or create an account to invite AI Trainers to your job.

LoginCreate account
Invite to Job
K

Ketevan Buskivadze

AI Training Data Specialist | Creative Writing, Finance & SFT Annotation

Georgia flagtbilisi, Georgia
$10.00/hrIntermediate

Key Skills

Software

No software listed

Top Subject Matter

Creative Writing & Literary Craft
Finance & Capital Markets
Marketing & Brand Strategy

Top Data Types

TextText
DocumentDocument
ImageImage

Top Task Types

Text GenerationText Generation
Data CollectionData Collection
ClassificationClassification
RLHFRLHF
Prompt + Response Writing (SFT)Prompt + Response Writing (SFT)
Evaluation/RatingEvaluation/Rating
Text SummarizationText Summarization
Question AnsweringQuestion Answering
SegmentationSegmentation

Freelancer Overview

Finance Research Analyst with hands-on experience producing AI training data through AfterQuery (Y Combinator W25), where I built decision-ready intelligence briefs and structured datasets used to train and evaluate large language models on financial reasoning tasks. Over the contract, I produced 20+ briefs distilling complex market data into clear, model-ready content and benchmarked 25+ competitor strategies across 3 market segments, sharpening my eye for the kind of accuracy, structure, and nuance that high-quality training data demands. My background blends marketing analytics, brand strategy, and finance coursework at Minerva University, which means I bring both quantitative rigor and strong written communication to annotation work. I'm comfortable working with ambiguous prompts, building rubrics from scratch, and producing outputs that are factually tight, well-reasoned, and free of the generic phrasing that weakens model training. I'm detail-oriented, fast, and used to working asynchronously across time zones.

IntermediateEnglish

Labeling Experience

Academic Integrity & Proctoring AI Training Data

ImageData Collection
Contributed image data for training AI proctoring systems used to detect suspicious behavior during online exams. Submissions captured a range of realistic test-taking scenarios and behavioral states, providing ground-truth examples that train models to distinguish between normal exam-taking activity and patterns flagged as potentially suspicious (eye movement, posture shifts, off-screen glances, environmental anomalies). The work supports the development of academic integrity tools that need diverse, naturalistic human data to reduce false positives and improve detection accuracy across different test-takers, lighting conditions, and devices. Quality measures included adherence to capture guidelines around framing, lighting, duration, and behavioral specifications for each labeled scenario.

Contributed image data for training AI proctoring systems used to detect suspicious behavior during online exams. Submissions captured a range of realistic test-taking scenarios and behavioral states, providing ground-truth examples that train models to distinguish between normal exam-taking activity and patterns flagged as potentially suspicious (eye movement, posture shifts, off-screen glances, environmental anomalies). The work supports the development of academic integrity tools that need diverse, naturalistic human data to reduce false positives and improve detection accuracy across different test-takers, lighting conditions, and devices. Quality measures included adherence to capture guidelines around framing, lighting, duration, and behavioral specifications for each labeled scenario.

2026 - 2026

Humanization Editing for LLM Outputs

TextRLHF
Edit LLM-generated drafts to remove characteristic AI tells, including overuse of em-dashes, formulaic transitions, sanitized vocabulary, predictable closing lines, and overly tidy emotional arcs, while preserving the original analytical or editorial substance. Work requires fluency in identifying what makes prose read as human (mid-thought pivots, sentence fragments, self-correction, register shifts, controlled imperfection) versus model-flavored (hedging, scaffolding, false balance, generic metaphor). Outputs serve as preference data and reference examples for training models to produce more natural-sounding text. Adhere to project-specific style guides covering vocabulary, punctuation, and voice consistency.

Edit LLM-generated drafts to remove characteristic AI tells, including overuse of em-dashes, formulaic transitions, sanitized vocabulary, predictable closing lines, and overly tidy emotional arcs, while preserving the original analytical or editorial substance. Work requires fluency in identifying what makes prose read as human (mid-thought pivots, sentence fragments, self-correction, register shifts, controlled imperfection) versus model-flavored (hedging, scaffolding, false balance, generic metaphor). Outputs serve as preference data and reference examples for training models to produce more natural-sounding text. Adhere to project-specific style guides covering vocabulary, punctuation, and voice consistency.

2026 - 2026

Creative Writing & Voice Workshopping

TextPrompt Response Writing SFT
Produce multi-turn human–AI conversations for AfterQuery's Shakespeare project, training LLMs on literary craft, voice, and creative reasoning. Each task spans 10–12 turns and requires writing both the human-side prompts (in persona as a poet, playwright, or other writer working through a specific craft problem) and humanizing the model-side responses to remove AI-tells while preserving substantive editorial reasoning. Tasks have included poetry workshopping (odes, haiku, tanka sequences), line-editing and copy polishing for theatrical scripts, and voice-and-craft critique. Work is governed by detailed style specs covering grammar register (clean vs. messy), word-count constraints, edit-category requirements (prose surface, narrator distance, structural reordering), and strict guidelines on avoiding chatbot phrasing, em-dashes, and overly tidy resolutions. Multiple tasks have passed final QA review.

Produce multi-turn human–AI conversations for AfterQuery's Shakespeare project, training LLMs on literary craft, voice, and creative reasoning. Each task spans 10–12 turns and requires writing both the human-side prompts (in persona as a poet, playwright, or other writer working through a specific craft problem) and humanizing the model-side responses to remove AI-tells while preserving substantive editorial reasoning. Tasks have included poetry workshopping (odes, haiku, tanka sequences), line-editing and copy polishing for theatrical scripts, and voice-and-craft critique. Work is governed by detailed style specs covering grammar register (clean vs. messy), word-count constraints, edit-category requirements (prose surface, narrator distance, structural reordering), and strict guidelines on avoiding chatbot phrasing, em-dashes, and overly tidy resolutions. Multiple tasks have passed final QA review.

2026 - 2026

Competitive Benchmarking Dataset Construction

TextClassification
Built a structured competitive landscape dataset covering 25+ companies across 3 market segments as part of contract work with AfterQuery (Y Combinator W25). Tagged each entry with positioning attributes, strategic moves, and market signals using a standardized schema designed for downstream LLM training. The work required strong consistency across hundreds of fields and frequent judgment calls — disambiguating overlapping categories, handling edge cases, and resolving conflicting source data. Quality measures included schema-adherence checks, cross-entry consistency reviews, and source verification for every tagged attribute. The dataset was delivered as a structured competitive landscape report and surfaced positioning gaps that informed downstream stakeholder decisions.

Built a structured competitive landscape dataset covering 25+ companies across 3 market segments as part of contract work with AfterQuery (Y Combinator W25). Tagged each entry with positioning attributes, strategic moves, and market signals using a standardized schema designed for downstream LLM training. The work required strong consistency across hundreds of fields and frequent judgment calls — disambiguating overlapping categories, handling edge cases, and resolving conflicting source data. Quality measures included schema-adherence checks, cross-entry consistency reviews, and source verification for every tagged attribute. The dataset was delivered as a structured competitive landscape report and surfaced positioning gaps that informed downstream stakeholder decisions.

2025 - 2026

Financial Intelligence Brief Generation

TextData Collection
Produced 20+ long-form financial intelligence briefs as training data for finance-focused LLM applications through AfterQuery (Y Combinator W25). Each brief required synthesizing earnings reports, market data, and competitor signals into a structured format with consistent reasoning chains, factual grounding, and traceable citations. Worked against detailed annotation guidelines covering tone, formatting, and source attribution to ensure outputs were model-ready and free of hallucinations or generic phrasing. Quality measures included peer review checkpoints, a factual-accuracy rubric tied to primary sources, and consistency audits across the brief set. Output supported stakeholders managing $2M+ portfolios, so accuracy and structural consistency were non-negotiable. Built strong fluency in producing dense, reasoning-heavy text annotations at scale.

Produced 20+ long-form financial intelligence briefs as training data for finance-focused LLM applications through AfterQuery (Y Combinator W25). Each brief required synthesizing earnings reports, market data, and competitor signals into a structured format with consistent reasoning chains, factual grounding, and traceable citations. Worked against detailed annotation guidelines covering tone, formatting, and source attribution to ensure outputs were model-ready and free of hallucinations or generic phrasing. Quality measures included peer review checkpoints, a factual-accuracy rubric tied to primary sources, and consistency audits across the brief set. Output supported stakeholders managing $2M+ portfolios, so accuracy and structural consistency were non-negotiable. Built strong fluency in producing dense, reasoning-heavy text annotations at scale.

2025 - 2025

Education

M

Minerva University

Bachelor of Science, Business – Strategic Finance, Economics, and Brand Management

Bachelor of Science
2023 - 2027

Work History

T

TUMO

Communications Intern

Buenos Aires
2025 - Present
M

Minerva University

Marketing & Communications Intern

San Francisco
2023 - Present