Product Work | Allison Londerée

Early Indicators of Student Success

Wed, 15 Apr 2026 12:00:00 -0400

At PERTS, I contributed to research on whether classroom learning conditions could function as early, actionable indicators of later academic outcomes. The question was simple but important: if we can measure whether students feel supported, challenged, and meaningfully engaged in class, can that help educators respond before end-of-term outcomes make problems obvious?

Figure note: Chance of earning a B or better by student group under negative (<5) versus positive (>=5) learning conditions on a seven-point composite scale, including race and reduced-price lunch status.

Highlights

Learning conditions are early indicators, not just “soft” context.
Better classroom conditions are strongly associated with better math outcomes.
This supports practical, repeated measurement so educators can intervene sooner.

Problem

Schools often rely on lagging indicators such as final grades or end-of-year test scores. Those measures matter, but they arrive late. Educators need earlier signals that point to whether students are experiencing the kinds of classroom conditions that support learning.

This project focused on three conditions in math classrooms:

Teacher Caring
Meaningful Work
Feedback for Growth

Approach

We analyzed whether these learning conditions predicted later math performance and whether changes in those conditions over time were associated with changes in student outcomes. The broader goal was to support a continuous-improvement approach in which educators could measure conditions, respond, and reassess rather than waiting for outcomes after the fact.

Methods

The report draws on data collected with the Character Lab Research Network from more than 4,000 U.S. students in grades 8 through 12 during the 2019-20 school year. Students rated learning conditions on a seven-point Likert scale, and those measures were linked to math grades over time.

Analyses examined:

the relationship between learning-condition ratings and the likelihood of earning a B or better in math
whether the same relationships held when controlling for demographics and prior grades
whether changes in learning conditions between October and February predicted later changes in achievement
whether results differed across student groups, including students eligible for Free and Reduced Price Lunch

Outcome / Impact

The findings make a strong case for treating learning conditions as decision-useful signals rather than soft background context.

Students who rated learning conditions most positively were more than twice as likely to earn a B or better in math.
Each step up in the composite learning-conditions score was associated with about 6% more students earning A or B grades.
A positive two-point shift in learning conditions was associated with roughly a 17% higher likelihood of earning a B or better in the following term.
Positive learning conditions were especially meaningful for students who had been less well served, including students eligible for FRPL and Black students.

What I like most about this work is that it connects careful behavioral measurement to a concrete intervention model. It is not just an explanatory report; it points toward a system educators can actually use to monitor conditions and improve them over time.

Tools

Education research, survey measurement, longitudinal analysis, regression-style modeling, applied behavioral science, research communication.

Next Mission Group Chrome Extension

Wed, 15 Apr 2026 12:00:00 -0400

I built an MVP Chrome extension for Next Mission Group around a concrete product problem: making military experience legible to civilian hiring workflows in real time—not as a one-off résumé rewrite, but in the browser, on the same pages where veterans already search and apply.

The work combined product flow, ontology design, heuristic scoring, LLM-assisted narrative, and production auth/API choices (Manifest V3, JWT connect flow, secrets kept off the client).

Highlights

Manifest V3 extension with a side panel UX: sign in, maintain a profile, open a supported job posting, run a fit analysis, save results, and leave feedback.
Job-board coverage for the MVP path: detection and parsing on LinkedIn, USAJOBS, Workday, Greenhouse, and Lever (with a pattern for adding more hosts).
Ontology-backed mapping between military roles, civilian occupations, and transferable skills—separating a canonical backbone from an inferred layer so LLM suggestions can be reviewed and traced.
Backend on Wix Velo for the MVP operational path: member auth integration, server-side API routes, and Wix Secrets (or env vars locally) so provider keys never ship in the extension bundle.

Problem

Military experience often does not align with how civilian roles are written: different vocabulary, credential expectations, and occupational categories. Veterans may be strong fits for a posting while still looking “off-profile” to both applicants and automated systems. The goal was to shrink that gap at the moment of job evaluation—not only after the fact on a separate site.

User flow

Install the extension and open the side panel.
Connect via the short-lived JWT flow tied to Wix Members (extension behaves like a first-party client without embedding API secrets).
Sync or edit profile data used for matching (roles, skills, history—details governed by product spec).
Navigate to a supported job posting; the extension detects the page type and extracts structured signals where possible.
Request an analysis: ontology + scoring produce an interpretable match; LLM copy is used where it adds explanation, with guardrails and provenance in mind.
Save outputs and optionally submit feedback to improve mappings and messaging over time.

Architecture (high level)

Layer	Role
Extension (MV3)	Side panel UI, content scripts / service worker per MV3 constraints, calls to Velo HTTP functions
Wix Velo	Authenticated APIs, persistence, orchestration; secrets for third-party / LLM providers
Ontology	Canonical identifiers + reviewable inferred links; scoring and narrative sit on top
LLM	Explanations and assistive text where helpful—not a black-box replacement for structured mapping

Security and operations

No long-lived API keys in the browser; privileged calls go through your backend.
JWT-based connect aligns the extension with the same membership model as the main site.
Provider keys live in Wix Secrets (or local env for development), consistent with a path toward harderening for review / pilot.

Outcomes

The MVP establishes an end-to-end path: install → connect → profile → analyze → save → feedback, with documentation for ontology growth, human review of inferred edges, and broader military–civilian coverage over time. Quantitative pilot metrics (activation, analyses per user, qualitative hiring outcomes) can be layered in as releases move beyond internal and friendly-user testing.

Demo

MP4 (recommended) — Put your file in this same folder as index.md, e.g. content/work/next-mission-group-extension/demo.mp4. Optional poster image: demo.jpg (same base name; used as the thumbnail before play).

You can use any filename: change src above to match. Files in assets/media/ also work (use a path like my-folder/recording.mp4). A direct HTTPS link to an .mp4 works as src too.

YouTube — If you prefer, set demo_youtube_id in the front matter (the v= ID only) and uncomment the line below; remove or comment out the video shortcode above to avoid two players.

GitHub

Your public links entry points to . When the extension (or a mono-repo) is published under your account or an org, add a second code link in front matter with that repository URL so the header shows Extension source alongside GitHub profile.

Tools

JavaScript, Chrome Extension APIs (Manifest V3), Wix Velo, JWT auth, ontology modeling, heuristic scoring, LLM integration, product and architecture documentation.

Pesticide Risk & Respiratory Health Modeling

Wed, 15 Apr 2026 12:00:00 -0400

As part of a Spring 2026 team project, I worked on a county-level modeling pipeline that linked pesticide exposure estimates to respiratory health burden in the United States. The goal was to build a planning-oriented model that could help public health stakeholders identify where asthma and COPD burden may warrant closer attention—while explicitly accounting for the ethical and equity implications of doing so.

Highlights

Built county-level XGBoost models to estimate relative asthma/COPD burden tied to pesticide-related context.
Published an interactive map for planning and resource prioritization, not diagnosis.
Paired performance metrics with equity/error analysis to make limits and risks explicit.

Problem

Agricultural pesticide use varies substantially by crop and region, and exposure risk is not evenly distributed. These differences often intersect with existing structural inequities in environmental exposure and healthcare access.

Public health agencies, health systems, and insurers need tools to prioritize prevention and outreach, especially when resources are limited. However, tools that surface “high-risk” areas can unintentionally reinforce stigma or obscure underlying structural drivers if not carefully designed and interpreted.

This project asked a practical question: can we combine public exposure, land-use, and health data to flag counties where pesticide-related respiratory burden may be higher—while maintaining transparency about uncertainty and equity impacts?

Approach

We built a county-year modeling pipeline for 2018 and 2019 that joined pesticide use estimates with respiratory-health outcomes and county context. The final product included both a documented modeling workflow and a public-facing GitHub Pages site with an interactive county risk map.

From the outset, the work was framed as a population-level planning tool, not a diagnostic or causal system. We paired model development with a “justification-for-proceeding” framework to surface assumptions, risks, and potential downstream harms before they became embedded in the system.

This included:

defining appropriate use (resource prioritization, not individual prediction)
identifying potential misuse (e.g., stigmatizing communities or over-attributing causality)
evaluating how model errors might differentially impact populations

Methods

The modeling dataset combined:

CDC PLACES county-level indicators, including asthma, COPD, smoking, obesity, and diabetes
USGS / EPA pesticide use estimates
USDA Cropland Data Layer features
ACS demographic covariates

The final selected model was an XGBoost regressor using a feature set that included 445 pesticide mass features plus baseline demographic, health, cropland, and time covariates.

In addition to standard modeling steps (exploratory analysis, spatial train/test splitting, hypothesis testing, and holdout validation), we conducted equity-focused evaluation:

compared model error across demographic and geographic subgroups
examined whether prediction accuracy varied systematically in counties with different socioeconomic or population characteristics
assessed where data sparsity or measurement limitations could bias results

The project also included model-card style documentation to make assumptions, limitations, and intended use explicit.

Outcome / Impact

The final XGBoost models performed strongly on held-out county-level validation data:

Asthma (CASTHMA): R² = 0.835, RMSE = 0.389, n = 1219
COPD: R² = 0.885, RMSE = 0.765, n = 1219

The project also produced a live interactive map that translates model output into a more usable planning interface for non-technical stakeholders.

Importantly, the equity analysis surfaced a meaningful gap:

model performance was not uniform across counties, with higher error observed in some regions with different demographic compositions and data coverage
this suggests that the areas most affected by environmental and health inequities may also be those where predictions are less reliable

This finding shaped how we positioned the tool:

as a starting point for investigation, not a definitive ranking
as a way to guide additional data collection and local validation, especially in underrepresented areas

More broadly, the project reinforced that building useful models in public health requires more than predictive performance. It requires clear communication about:

what the model can and cannot say
where it may fail
and how its outputs could impact real communities if used uncritically

Tools

Python, Jupyter, XGBoost, county-level public health data, geospatial joins, GitHub Pages, model-card documentation, responsible AI frameworks.

RSV Vaccine Attitudes in a Digital Health Intervention

Wed, 15 Apr 2026 12:00:00 -0400

This case study is based on published work from my time at Lirio on a digital health intervention designed to increase RSV vaccination among adults over 60. Rather than treating outreach as a one-way campaign, the project looked closely at what people actually texted back and used those replies to better understand attitudes, friction points, and behavior-change opportunities.

Highlights

Analyzed unsolicited SMS replies to understand RSV vaccine attitudes at scale.
Used structural topic modeling plus expert thematic coding for interpretable insights.
Findings informed how live intervention messaging could be improved over time.

Problem

When the RSV vaccine was introduced for older adults in the United States in 2023, many of the drivers of uptake were still unknown.

For teams designing outreach interventions, this creates a familiar challenge: how do you improve messaging when beliefs, concerns, and barriers are still emerging in real time?

Rather than relying solely on structured surveys or downstream outcomes, we turned to a more immediate signal: what people actually texted back.

Approach

The intervention was launched with a large community pharmacy chain using SMS precision nudging to encourage vaccination. As patients replied, we accumulated a large, messy, and highly valuable dataset of unsolicited responses.

However, the scale of the data introduced a new constraint: there were far too many replies for behavioral experts to review manually, and insights were needed quickly to inform an active intervention.

My role as the data scientist was to bridge that gap—developing a scalable way to extract structure and meaning from free-text responses, while preserving the nuance required for behavioral interpretation.

Methods

The study population included 2,481,987 eligible adults aged 60 and older, of whom 35,716 people replied, generating 105,848 text responses. After removing operational texts such as STOP and HELP, the analytic dataset included 46,964 unsolicited replies.

We used a mixed-methods workflow that combined computational modeling with expert interpretation:

Structural topic modeling (STM) to identify latent themes across the full dataset
Thematic analysis on a hand-coded subset, conducted by behavioral health experts
Alignment of model-derived topics with expert-reviewed themes to ensure interpretability
Characterization of topics along dimensions such as sentiment and practical vs. emotional function
Comparisons across time, message condition, and available covariates (e.g., flu vaccination history, insurance type)

This approach allowed us to scale insight generation without losing the behavioral context needed for intervention design.

Topic Map of 30 topics in text message replies to a precision nudging digital health intervention for RSV vaccination. Shape and colour indicate the mapping of structural topic model topics onto the 10 thematic analysis topics. Black star = Wrong recipient, teal circle = Help, red cross = Stop, purple plus = Already vaccinated, brown square = Will not get vaccinated, yellow triangle = Benign, and grey diamond = Nonsense. Size indicates topic proportions. Topics 1-10 are largest, topics 11-20 are of middle size, and topics 21-30 are smallest.

Outcome / Impact

The mixed-method approach of modeling and expert review converged to produce a grounded, interpretable map of how adults responded to a newly introduced vaccine in a real-world setting.

Key findings included:

expressed attitudes became less negative later in the intervention
individuals without prior flu vaccination and those with commercial insurance were more likely to express refusal
messages framed around emotional consequences generated the highest engagement
messages framed around anticipated regret generated the lowest engagement

Beyond the publication, this project demonstrated a practical pattern: behavioral data generated during live interventions can be rapidly structured and fed back into design decisions.

Instead of treating outreach as static, this approach enables a tighter loop between what people say, how we interpret it, and how interventions evolve.

Tools

Behavioral science, thematic analysis, structural topic modelling, text analysis, digital health experimentation, intervention evaluation.