Cadler D'ahiti
Data Science and Computer Science student at the University of Wisconsin–Milwaukee. McNair Scholar, NSBE Programs Chair, and the person behind Telio.
- Studying
- B.S. Data Science and B.S. Computer Science
- Graduating
- May 2028 (anticipated)
- Honors
- Dean's List since Fall 2024, Lawton Scholar, Patricia H. Weisberg Scholar, Project Lead the Way Scholar
- Writes code in
- Python, Java, R, SQL, VBA
- Speaks
- English, French, Creole
About me
I'm a student at UW–Milwaukee working toward two degrees, one in Data Science and one in Computer Science. I like work that sits between software and scholarship: I write the code a question needs, then study what it shows.
Outside class I run events for our NSBE chapter as Programs Chair and work in the Dean of Students office. Through the McNair Scholars Program I research how AI can use less energy, mentored by Dr. Martin.
Experience
Programs Chair, NSBE
I manage events for the chapter, and handle attendance reports, fundraising and networking for collaboration events.
Office Assistant, Dean of Students
Administrative support for the office's daily work, including confidential student records and communications.
Tour Guide, TRIO
Led prospective students and families around campus and answered questions about academic programs and campus life.
Member, Leaders Igniting Transformation
A student organization for political awareness. We encouraged students to vote and follow the news.
Sales Associate, Cornelier Fireworks
Helped more than 100 customers choose fireworks, managed inventory and processed transactions.
Projects
Telio
Scores prompts on readability, vocabulary variety and hedging, then logs real token usage and estimates energy cost.
Study AI agent
The agent participants use in the experimental part of my study.
Survey instrument
The questionnaire that goes with the agent.
Shareholder profit study
Gathered daily financial data for securities identified by CUSIP to see whether shareholders were making or losing money on them.
Lawn service manager
An object-oriented system for service requests, pricing calculations and report summaries.
Research deep dive: Telio and SCEF
Watch the presentationRead the bibliography
My McNair proposal, "Energy Efficient AI & Ethical Studies" (June 2026), argues that wordy, poorly structured input makes AI cost more energy than it needs to. Telio is the tool I built to test that.
The question
How do data structure and agent design choices jointly shape the energy cost and ethical reliability of AI systems?
Existing research treats these as separate conversations. Energy researchers study data centers, language researchers study output quality, and ethicists study responsible use. No one connects the three, so a well-meaning fix in one area, such as better prompts or more autonomous agents, can quietly raise costs in another.
Most talk about AI's energy use is about hardware: chips, data centers, cooling. My proposal looks earlier in the chain. It asks how much of the energy bill is decided by the quality of what people feed the model and by the choices developers make when they build around it.
The literature review links linguistic inefficiency to energy inefficiency, and points to a gap: there is little structured, replicable process for developing AI productively.
The framework: SCEF
The Structural–Computational Efficiency Framework is my own model. Its claim is that energy use is a downstream result of structure.
Structure
How well the input is written and organized, and how the system is designed.
Computation
The work the model has to do to handle that input.
Energy
What that work costs to run.
| E | Energy consumed |
|---|---|
| C | Computational load |
| S | Structural entropy of the input data, meaning how disorganized it is |
| A | Developer and agent design choices |
| H, T | Hardware efficiency and task complexity, held constant as controls |
Hypotheses
H1
Better-structured input leads to lower computational load, holding task difficulty constant.
H2
Lower computational load leads to lower energy consumption, holding hardware constant.
H3
Agent designs with explicit guardrails use less energy and are more accurate.
H4
The link between structure and energy holds across architectures, but its strength varies with hardware and task complexity.
What the data should show
If the framework holds, I expect three things.
- Better-structured prompts cost less. They lead to measurably lower token counts and energy estimates for responses of the same quality. That gives developers a low-cost way to cut energy use without new hardware.
- The savings level off. The energy-savings curve flattens as structure improves. Structure helps up to a baseline and is not a cure-all after that.
- Efficiency and accuracy improve together. This challenges the idea that the two trade off, which matters for ethical AI use and not only for cost.
The instrument: how Telio works
Telio is a Python program. For every prompt it runs the same four steps.
- Score the wording. Three measures: readability (Flesch Reading Ease), redundancy (type-token ratio, or how varied the vocabulary is) and ambiguity (hedge-word density, or how much of it is "maybe" and "kind of").
- Send the prompt. The prompt goes to the Claude API as written.
- Log the real cost. Telio records the actual token usage that comes back.
- Estimate the energy. Tokens are converted to an energy figure using a configurable constant taken from cited sources.
What Telio measures
| Measure | What it captures | Maps to |
|---|---|---|
| Flesch Reading Ease | Sentence-level readability and complexity | Proxy for S |
| Type-token ratio | Lexical diversity and redundancy | Proxy for S |
| Hedge-word density | Ambiguity | Proxy for S |
| Real token usage | Actual computational load, logged through the Claude API | C |
| Energy conversion estimate | Tokens converted to an estimated energy draw | E |
Try the scoring step
A simplified browser version of step one. It does not call the API or estimate energy.
Study design
Qualitative, experimental. Participants work with an AI agent I developed, then answer a survey I designed alongside it.
Quantitative, correlational. An analysis of educational data and energy consumption data, looking at how the two move together.
Controls and limitations
Controlled for
- Hardware efficiency (H). Chip generation and cooling infrastructure.
- Task complexity (T). Harder queries need more computation regardless of structure.
Open limitations
- Flesch bias. Flesch Reading Ease marks down academic and technical vocabulary, so it can undercount the quality of a complex but well-structured prompt. I disclose this as a limitation.
- Estimated energy. The energy and carbon conversion constants come from the literature. They are not measured hardware output.
- One tool. Findings from a single custom tool may not generalize to every model architecture or provider without replication.
What the literature says, and what it leaves out
I read ten sources for the proposal. They agree that AI's energy use is large and growing. None of them measures energy at the level of a single prompt, which is the gap my study works in.
The scale of the problem
- Kandemir (2025), Penn State Institute of Energy and the Environment. Data centers used 4.4% of U.S. electricity in 2023, and energy use by GPU-accelerated AI servers grew from under 2 TWh in 2017 to more than 40 TWh in 2023. It does not look at how prompt quality affects energy use.
- Jonker & Gomstyn (n.d.), IBM Think. Explains what sets an AI data center apart and reports a projected 165% rise in data center electricity demand by 2030. It covers hardware only.
- Eilam et al. (2024), IEEE Transactions on Semiconductor Manufacturing. Specialized hardware cuts operating energy but raises the emissions built into manufacturing. It offers a framework and no new measurements.
- Mahmood et al. (2025), NTT Technical Review. Proposes moving data center workloads to where and when renewable power is available. Japan curtailed more than 2 billion kWh of renewable output in fiscal 2024.
- Berenjian et al. (2021), Journal of Intelligent & Fuzzy Systems. A task scheduling algorithm that cut energy use by up to 58.3% in simulated cloud data centers.
Agents and the structure of their input
- Gutiérrez Tafoya et al. (2026), International Journal of Combinatorial Optimization Problems and Informatics. The closest study to mine. Restructuring one source document improved a conversational agent's answers on most quality measures. It did not measure whether better structure also lowers computing or energy cost.
- Dautenhahn (1998), Applied Artificial Intelligence. A foundational framework for designing socially intelligent agents. Written long before anyone asked what such agents cost to run at scale.
- Hinov (2026), AI. A reference architecture for AI agents that run power grids under human supervision. It does not count the energy the agents themselves use.
- Lan (2026), Beijing Review. Describes China's national push to deploy AI agents, and gives my introduction its "passive response" framing.
How people use AI
- Radday & Mervis (2026), Educational Leadership. Argues schools should teach students to apply AI to real problems instead of using it for quick answers. It draws on Brynjolfsson et al. (2025) for labor-market data.
Download the full annotated bibliography (PDF)
Where it stands
Proposal
Drafted in full: introduction, literature review, background, theoretical framework, assumptions, hypotheses, predictions and APA references.
Telio
Built and working, with a README and example prompts.
Results
Not yet. Data collection comes next, and the findings will be added here when it is done.
Presentation
14 slides and a speaking script, in three parts: Background, Research Structure, Further Steps. Watch the presentation on Prezi.