GPA: 4.00 UW / 5.361 W · Rank 11/481 · ACT 36 · AP Scholar with Distinction

AP exams: 9 taken, a 5 on all 9 — Computer Science A, Statistics, Physics 1, Seminar, English Language & Composition, U.S. History, World History, Psychology, Human Geography

Coursework: AP Calculus BC, AP Physics 2, AP Physics 1, AP Computer Science A, AP English Language, AP English Literature, AP Government, AP Economics, Computer Science III, Mobile App Programming, Spanish 2

Honors: IEEE-EMBS IECBES 2026 paper accepted for presentation · X-Step startup: Conrad Innovator 2025–2026 — Health & Nutrition; Texas A&M McFerrin Center for Entrepreneurship Texas Ideas Challenge — Top 40 Finalist; Diamond Challenge 2026 — Semifinalist · ShopAI (as ShopSmart): TechLit Bridging the Gap Hackathon — 1st Place (Aug. 2025); PeddieHacks — Best Innovation Award, 1st/250 (Aug. 2025); Hack4Humanity — 3rd Place, 3/500 (Oct. 2025); AsterHacks — 5th Place, 5/250 (Oct. 2025)

Mentor: Joe Xiao (PhD student) · Under peer review, Frontiers in Aging Neuroscience — submitted August 04, 2026 · Preprint medrxiv.org/content/10.64898/2026.08.04.26359730v1

• Question: can a patient's speech across clinic visits forecast their next Mini-Mental State Exam (MMSE) score? Framed as supervised regression on the DementiaBank Pitt Corpus (173 longitudinal samples, 126 patients) using only earlier visits.
• Representation: to capture how discourse organization changes with decline, extracted 12 per-visit linguistic features (POS fragment ratio, named-entity rate, 50-token MATTR, filler rate, perseveration) and Sentence-BERT coherence features (adjacent-sentence cosine similarity, global coherence, semantic slope, topic drift, embedding variance), plus visit-to-visit Δ features.
• Model: reduced 384-d SBERT embeddings to 20 PCA components, added 9 clinical/demographic covariates (53 dimensions), and tuned LightGBM with Optuna/TPE over 200 trials (RobustScaler, correlation pruning, permutation-importance selection).
• Evaluation: found patient-level leakage (one patient on both sides of a split) and rebuilt evaluation with patient-grouped 5-fold GroupKFold, augmented copies kept in their patient's fold, original samples only scored: R² 0.840 ± 0.015, RMSE 2.75, MAE 2.02 MMSE points. Ablation: clinical history carries most of the signal (clinical-only R² 0.72, SBERT-only 0.46, linguistic-only 0.12); speech adds a smaller gain (0.76 combined); non-nested tuning is disclosed as an optimism limit. Code github.com/ddevatha1/dementia-mmse-prediction

Accepted for Presentation — IEEE-EMBS IECBES 2026 · Kuala Lumpur, Dec. 9–11, 2026

UC Santa Cruz AIEA Lab (Prof. Leilani Gilpin) aiea-lab.github.io · deeptanshudevatha.com/research/x-step

• Question: can four strategically placed sensing sites capture regional pressure trends comparable to a 32-sensor insole? I engineered the machine-learning component: a 59-feature representation of the team’s four-site plantar-pressure readings (per-site peak/mean/CV, pressure–time integral, cadence, stance ratio, anterior load share) and a 540-parameter logistic-regression framework that classifies biomechanical risk.
• Designed the correlation study (149 recordings, 15 adults, public 32-cell insole): four sites keep waveform shape (median r = 0.72–0.99) but not magnitude (peak-to-peak r = 0.43–0.74); center-of-pressure tracking keeps anteroposterior motion (r = 0.81) but loses mediolateral (r = 0.01).
• Quantified subject-identity leakage on a synthetic 24-subject cohort (macro-F1 0.885 subject-grouped vs 0.931 ungrouped) and documented the generalization limit.

• Prices differ by store and no retailer offers an API built for comparison, so ShopAI searches nine chains at once and plans the trip. Built end to end: React Native/Expo client and an Express/TypeScript backend on AWS EC2 behind Nginx, with adapters that turn unlike sources (OAuth2 REST, cookie-session GraphQL, Playwright, server-rendered HTML) into one product schema. Live on the App Store. apps.apple.com/app/shopai/id6797083809
• Concurrent fan-out search under a shared response budget with per-store failure states, query-correction/relevance/ranking, and a per-ZIP auth warm-up state machine. A deterministic planner enumerates all 127 non-empty subsets of the 7 live-data stores (exhaustive search is cheap, so no heuristic) for cheapest, fastest, fewest-stop and weighted plans, with zero LLM calls. deeptanshudevatha.com/projects/shopai

• Founded a 3-year Python workshop serving 124 students across 10 Frisco neighborhoods; recruited peers to help teach. Wrote the curriculum, 120 one-hour classes (5 per week for 8 weeks each summer) and 20 projects: control flow, file I/O, JSON, exceptions, OOP/inheritance and assert-based tests, then spaCy NLP (tokenization, lemmatization, POS tagging, sentence segmentation, NER, Matcher patterns, a rule-based chatbot). deeptanshudevatha.com/projects/kidcode
• To test learning, wrote 13 pre/post assessment pairs (same skills, different code): across 104 students the mean score rose 41.7% → 58.8% (+17.1 pp; 99 improved, 4 unchanged, 1 declined; no comparison group). Students completed 1,214 working programs, 572 using spaCy. deeptanshudevatha.com/projects/kidcode/assessment

• X-Step is a student startup building a smart insole for people at risk of diabetic foot ulcers, which form where pressure builds up unnoticed. The team's ESP32 insole streams four pressure sensors over BLE; I wrote the software: the React Native/Expo app (BLE client, live pressure heatmap, daily risk score, alerts, trends, guided routine) and the earlier Flask/SQLite + LLM health-assistant backend. deeptanshudevatha.com/projects/x-step
• Trained two CNN prototypes on public data: EfficientNet-B2 on 9 foot-pressure heatmap postures (98.3%, random 80/20 split) and ResNet-50 on four wound-severity grades. Intended flow: an ulcer photo's severity score plus care habits go to an LLM that drafts routines. Bench-stage prototype; no wearer data, no clinical validation.

• Taught 38 weekly Zoom sessions of Python and math to 11 children at the Ayusha Nilayam orphanage in the Alwal/Kompally area of Hyderabad, adapting to one shared laptop, unreliable internet and a Telugu–English language gap; then two weeks in person. deeptanshudevatha.com/projects/ayusha-nilayam

Dr. Chris Davis · Built a rule-based text simplifier: dependency parsing, lexical substitution, clause splitting, passive-to-active rewriting.