Sandra B.

Key Skills

Software

Other

Top Subject Matter

Education research and faculty training on generative AI use

Instructional AI systems for persistent career-education workflows

Responsible AI implementation with evidence labeling for professional domains

Top Data Types

Text

Document

Top Task Types

Prompt + Response Writing (SFT)

Function Calling

Entity (NER) Classification

Data Collection

Freelancer Overview

I work on the human side of AI training data: collecting it, structuring it, and evaluating model output against it. I built a self-hosted human-feedback platform on LimeSurvey with a custom SQL schema and a participant interface engineered down to response-order and contrast effects, owning the pipeline from instrument design through analysis in R and Python under IRB and data-residency constraints. Alongside that, I have spent two years developing domain-specific prompts and assessing LLM responses across Claude, ChatGPT, and Gemini, including an evidence-labeling rubric that forces each output to separate fact, inference, and unknown. My edge is domain judgment: I catch fluent answers that are quietly wrong, and I can explain the failure precisely enough to act on.

Labeling Experience

Quantitative Data Collection & Human-Data Research Platform — survey platform engineering (2026)

OtherDocumentData Collection

Engineered a self-hosted human-feedback survey platform to support end-to-end collection and ownership of IRB-ready raw data for research. Implemented a custom database schema and rebuilt the participant interface to treat response-order and presentation effects as methodological variables. Enabled unlimited instruments and analytics access for R and Python while retaining full raw-data ownership for data-residency requirements. • Built on open-source LimeSurvey to own the feedback data pipeline. • Secured the application layer and created a custom SQL schema for survey data. • Reconstructed the participant interface with controlled typography, contrast, and response-order effects. • Provided direct database access and analytics support for downstream research workflows in R and Python.

2026 - 2026

Beneficial by Design — production-grade open-source AI implementation system with evidence labeling (2026)

OtherTextEntity (NER) Classification

Established an evidence-labeling discipline requiring AI outputs to distinguish facts, inferences, and unknowns before presenting results to end users. Authored auditable, profession-specific AI skill files and a research orchestrator combining multiple analytical frameworks and structured output templates. Produced practitioner documentation and implementation guides to operationalize the labeling discipline across multiple LLM platforms. • Defined and enforced fact/inference/unknown distinction requirements in AI output generation. • Built profession-specific auditable skill files to standardize evidence labeling. • Designed a research orchestrator using analytical frameworks with structured outputs. • Authored implementation handbook with guidance for Claude, ChatGPT, Gemini, Codex, and Claude Code.

2026 - 2026

Persistent AI Operating System (Claude Code + Obsidian) — design of a persistent, instruction-based AI operating system (2026)

OtherTextFunction Calling

Created and operationalized domain-calibrated instruction modules and persistent guidance mechanisms to structure AI behavior for career-education use cases. Implemented standards compliance as an automated, structural feature rather than manual oversight. Measured and optimized performance impacts including cost and prompt-engineering overhead reductions. • Engineered multi-module instruction sets for consistent LLM behavior in domain tasks. • Built an auto-memory layer that writes behavioral corrections to persistent files and reloads them at session open. • Enforced hook-based configuration via settings.json to reduce session context loss. • Optimized startup token costs and reduced prompt-engineering overhead through pre-loaded modules.

2025 - 2026

Reimagining Education with Claude (Faculty AI Program) — end-to-end instructional design and AI-enabled training (2025)

Don't discloseTextPrompt + Response Writing (SFT)

Led end-to-end instructional design for an asynchronous faculty AI program integrating Claude Projects into authentic teaching. Operationalized Constitutional AI principles with transparent guidance, harm reduction, and alignment with human intent across instructional tasks. Conducted mixed-method evaluation to measure AI-literacy gains resulting from the training. • Designed instructional modules and learning activities for LLM use in teaching contexts. • Ensured rigor, transparency, and professional agency in AI-enabled instructional workflows. • Applied Constitutional AI guidance to instructional task designs. • Evaluated outcomes via mixed-method assessment of learner AI literacy.

2025 - 2025

Education

Work History