The capstone project gives participants an opportunity to apply AI concepts, data reasoning, and ethical analysis to a meaningful, real-world problem. Rather than emphasizing complex coding, participants use AI assistants and low-code tools to support thoughtful problem framing while evaluating the opportunities and limitations of their approach.
Note: Unless otherwise noted, projects in this collection use publicly available, educational, or simulated datasets. The resulting analyses, prototypes, and visualizations reflect participant learning and exploration and may not represent fully developed or deployed solutions.
2025 – 2026 Capstone Projects
Cohort Stories
Bill Wittland
Summary
Bill Wittland’s capstone is a portfolio of seven publication-quality stories profiling the work of his fellow ExLENT participants: 7 profiles of other capstones (Cookie Monsters, Rolling Homes, Team A.H.A., Hay, Messaoudi, Powell) plus a first-person reflective piece on his own ExLENT journey. Drawing on his career as a professional marketer and writer, Bill conducted real interviews with each participant, captured original quotes, and produced consistent profile-format pieces with deliberate headlines (“Optimizing Delicious,” “Streamlining Safety,” “Assumptions Vs. Data”) and standardized two-paragraph program-context boilerplate that makes the stories deployable in funder reports, alumni newsletters, and recruitment materials with no rework. His self-reflective piece names the program’s central pedagogical insight — the move from AI-as-search to AI-as-maker — more cleanly than the program’s own materials typically do, and the work as a whole stands as the most program-validating capstone in the cohort: a non-STEM working professional turning his existing craft into a deliverable the program itself can use.
Optimizing Cookie Baking
The Cookie Monsters: Keli Bedics, Eric Derr & Alexa Luna
Summary
The Cookie Monsters built a linear-programming optimization tool for a small-batch bakery (Chips & Doodles, owned by team member Eric Derr’s mother) that takes a pantry inventory and ingredient prices as inputs and returns the optimal mix of three cookie recipes- chocolate chip, snickerdoodle, and oatmeal raisin- to either maximize total cookie output or hit a required production target with maximum extras. Built in Python with PuLP and deployed as an interactive HTML dashboard, the tool models 14 unique ingredients, integer batch decisions, and pantry-quantity constraints. Test runs validated the model against hand-calculated outputs and surfaced design improvements for future iterations. The team punctuated its presentation by baking and distributing the actual cookies to the class, turning the optimization output into something the audience could literally taste.
RV Wholesale Shipments and Retail Sales
Team Rolling Homes: Piper Burrows, Arlissa Lile, Laurel O’Brien & Kevin Ricksgers
Summary
The Rolling Homes investigated whether RV wholesale shipments can be used to predict retail sales, a real demand-forecasting question for a regional industry where production lead times create persistent volatility. Using monthly industry statistics, wholesale shipments, and retail sales data, the team built linear regression models in R, including time-lagged variants to capture delayed effects and time-series visualizations to show how the two series move together. Their analysis demonstrated a positive relationship between shipments and downstream sales, with lagged models capturing real-world supply-chain dynamics. The team paired the quantitative work with an explicit ethical reflection on correlation vs. causation, national-level limitations, and the human consequences when forecasts miss in a labor-reactive industry.
Manufacturing Knowledge Governance Chatbot
Team A.H.A.: Aidan Gibson, Aina Muça & Heather Winrotte
Summary
Team A.H.A. built a controlled, document-grounded knowledge system designed for industrial manufacturing environments, not a generic chatbot, but a governance-enforcing operator assistant. The system uses FAISS semantic vector search over chunked, versioned PDF work instructions, returns answers with explicit citations in the operator’s detected language (English or Spanish), and automatically escalates safety-related queries, expired documents, or permission-seeking phrasing (“Can I…”, “Is it okay to skip…”) to a supervisor rather than answering. A separate supervisor dashboard logs queries, flags soon-to-expire documents, and supports ISO 9001 and OSHA audit traceability. The project was selected by independent judges as the cohort’s 1st place capstone.
Material Percentage of Revenue
The Lucky Charms: Dylan Brayton, Skylar Jones & Leina Ulutoa
Summary
The Lucky Charms addressed a real corporate finance question: when material costs rise faster than revenue, what is actually driving the gap – volume, mix, or rate – and is the pressure concentrated at the part level or the customer level? The team built a variance-bridge analysis, decomposing year-over-year changes into volume, mix, and rate components, applied at both part and customer granularity. They first prototyped the framework on AI-generated proxy data (allowing them to validate the methodology before applying it to sensitive real data) and then translated the working framework to the actual dataset. Their headline finding from the sample data: the change was overwhelmingly volume-driven, with mix helping marginally and rate pressure offsetting that gain.
AI in Manufacturing Procurement at GSM
Ashley Hay (Purchasing Manager, General Stamping & Metalworks)
Summary
Ashley Hay’s project evaluates the potential use of Claude within a procurement function at General Stamping & Metalworks (GSM) – a metal stamping and fabrication manufacturer with operations in South Bend, Indiana and Tomah, Wisconsin. Hay’s analysis breaks down the manual administrative workload for the two-person purchasing team (email follow-ups, ERP updates in Plex, compliance-document tracking, due-date reconciliations) and identifies five concrete Claude-enabled functions: supplier email monitoring, email-vs.-ERP comparison, discrepancy flagging, automatic compliance-document linking to Plex, and a buyer-priority dashboard. The proposal is supported by a five-phase implementation plan (process mapping, opportunity identification, pilot, measurement, full rollout) and a rigorous risks-and-limitations section that frames Claude as decision support, not a buyer replacement.
Southwest Asia and North Africa (SWANA) Funding Architecture
Sarra Messaoudi
Summary
Sarra Messaoudi’s project applies network science to humanitarian, peace-building, and development funding flows across 13 SWANA countries, unifying four authoritative datasets (OCHA CBPF, OCHA FTS, OECD CRS, and UNDP MPTF) into a single 1,536-organization, $27.9B network. Messaoudi produced a live interactive dashboard along with four substantive findings: a funding-concentration Gini of 0.91 (the top 10 organizations receive 48.8% of all funds); a 3× localization gap between INGOs and local NNGOs against Grand Bargain commitments, unmet in 7 of 8 countries; 62 articulation-point organizations whose removal would disconnect parts of the network; and stark country variation (Iraq 7.9% local funding share vs. Somalia 41.4%). The work was contextualized with ACLED conflict scores and CIVICUS civic-space ratings.
Community Development Disparities of St. Joseph County
Jackson Powell
Summary
Jackson Powell’s capstone is a tract-level analysis of socioeconomic and racial disparities in St. Joseph County, Indiana, integrating five public data sources (ACS S1901 income, ACS DP05 demographics, county property records, the U.S. Census Bureau Batch Geocoder, and TIGER/Line tract shapefiles) into one neighborhood-level model. His analysis shows tract median household income ranges from $23,750 to $161,520 (a 6.8× gap), with a +0.61 correlation between income and % White and a −0.61 correlation between income and % Black. Property values mirror the same pattern. He frames the findings carefully, acknowledging dataset limitations and noting that statistics cannot tell the whole story, while landing firmly on the conclusion that the inequalities are systematically linked across race, income, and housing wealth, not random.
Indiana School Choice & Academic Outcomes
John Facchini
Summary
John Facchini’s capstone is an interactive four-panel HTML dashboard analyzing Indiana’s Choice Scholarship Program, now one of the largest school-voucher programs in the United States, against statewide ILEARN academic-outcome data. Drawing on the IDOE Annual Choice Report 2024–25 and ILEARN Spring 2025 statewide summaries, he documents the program’s scale (76,067 participants, $497M in scholarships, 373 schools, $6,536 average award against $8,369 average tuition) and its dramatic shift in purpose: after the 2022–2023 eligibility expansions, nearly 70% of current Choice participants have no record of ever attending an Indiana public school. The dashboard pairs this with a careful five-year ILEARN trend analysis showing modest, uneven recovery from pandemic-era disruption rather than transformation, and explicitly frames why direct public-vs.-Choice score comparisons are statistically unreliable without rigorous matching.