Services

AI data work, done with care

Six ways I help AI teams build models that are accurate, safe and pleasant to use. Each service draws on specific experience, explained below.

Core service

Response rating & ranking (RLHF)

Human feedback is how models learn what a good answer looks like. I compare responses side by side, score them on each rubric dimension, and explain my choice so reviewers can trust and reuse it.

Pairwise preference and multi-response ranking
Rubric scoring: accuracy, helpfulness, safety, tone, instruction-following
Short, specific justifications that cite the exact problem
Consistent calibration against reference examples

Draws on: following clinical protocols exactly, written reporting at Lopfatec, and writing a 41-item weighted rubric for an expert benchmark task.

Example deliverable

A rating with a rationale a reviewer can verify quickly:

Preference: B (much better) Accuracy A 2/5 · B 5/5 Safety A 1/5 · B 5/5 Reason: A advises massaging red skin over bony areas, which can worsen pressure damage. B gives the safe repositioning schedule and says when to call a nurse.

Try the full interactive sample

English writing quality review

A correct answer can still fail the user if it is stiff, wordy or confusing. I review and rewrite AI responses so they read naturally and get to the point.

Grammar, spelling, punctuation and consistency
Clarity: plain language, short sentences, logical order
Tone that fits the user's situation, warm but not over-familiar
Rewrites that keep the original meaning intact

Draws on: professional written communication with vendors and clients, and preparing reports and meeting notes.

Example deliverable

Before: "Food pantries is places where peoples can getting food when there budget are low." After: "A food pantry gives free groceries to people who are short on money for food." Notes: fixed agreement errors; simpler wording.

See the before/after sample

Health & caregiving review

People ask AI how to care for a parent, a child or themselves. I check that caregiving answers are accurate, practical and clear about when to get professional help.

Patient comfort, hygiene, mobility and pressure-injury prevention
Vital signs and when changes need attention
Infection control and wound-care basics
Clear "seek care now" guidance when warning signs appear

Draws on: bedside care for eight patients per shift at Parirenyatwa Hospital, and healthcare prerequisite coursework. I am not a licensed clinician and will say so where a clinical reviewer is needed.

What I check in every health answer

Is every instruction safe to follow at home?
Does it name the warning signs and when to get help?
Is it free of outdated or myth-based advice?
Is it written for a worried, non-medical reader?

Safety & sensitive topics

Some prompts come from people in difficult moments. I evaluate whether AI answers on crisis, hardship and wellbeing are safe, respectful and point to real help.

Flagging harmful, dismissive or overconfident advice, with severity
Checking that urgent situations get urgent guidance
Respectful handling of food insecurity, housing and social-services questions
Avoiding judgment, stereotypes and stigma

Draws on: social work coursework (person-in-environment, NASW Code of Ethics) and supporting students facing food insecurity.

Severity scale I use when flagging

High: could delay urgent care or cause harm if followed
Medium: misleading, incomplete or likely to confuse
Low: tone or clarity issues that weaken trust

See a flagged sample

Data labeling & QA

Models are only as good as their labels. I tag, classify and check data accurately and consistently, and I catch the errors others miss.

Text classification and tagging to a defined schema
Quality review of other annotators' work
Finding gaps, duplicates and inconsistencies
Careful handling of personal or sensitive data (PII)

Draws on: inventory audits and confidential intake records at the campus food pantry, and document classification at Lopfatec.

Example QA note

Item 214: label "billing" → should be "eligibility" (user asks whether they qualify, not about cost) Item 231: contains a phone number → mask as [PHONE] per privacy guideline 4.2

Prompt & answer writing

Good training data sounds like real people. I write realistic prompts across different users and situations, plus high-quality reference answers to match.

Realistic prompts in varied voices and levels of detail
Reference answers that follow the style guide exactly
Edge cases that test safety and instruction-following
Caregiving, wellbeing, education and everyday topics

Draws on: real conversations with patients, families and students, and authoring a full agentic task (scenario, brief, emails, deliverable spec) for an AI benchmark.

Example prompts

"my mum keeps forgetting to drink water since her stroke, any tricks that actually work?" "I'm a student and my aid got delayed. Where can I get food this week without paying?"

Working together

Engagement details

Remote & flexible

I work remotely from Massachusetts with flexible hours on U.S. Eastern time.

Project-based or ongoing

Short evaluation batches, or a longer role on a rating or writing team.

Guideline-driven

I follow your rubric, style guide and privacy rules exactly, and ask when something is unclear.

Need one of these services?

Tell me the task type, volume and timeline, and I'll reply with my availability.

Start a conversation See work samples