Xinheng Wang

Cross-platform surrogacy app & web admin console

A production system with mobile apps for surrogates and intended parents, plus a web admin for agency staff. I translated the founder's requirements into architecture, implementation, QA, and official Apple App Store / Google Play releases.

• iOS & Android

  Shipped with Expo / EAS. User journeys cover applications, match milestones, medical reports, IVF/OB appointments, and reimbursement claims.

• Web admin

  Next.js (TypeScript) on Vercel, with role-based access for Admin, Branch Manager, and Finance Manager (read-only payments).

• Data & auth

  Supabase PostgreSQL with RLS on the mobile client, server-side auth on the admin, and Realtime for match updates and notifications.

• Release

  EAS production builds for store binaries; Expo OTA for JavaScript-only fixes without a full store resubmission.

LLM interpretability, semantic surprise, and neurocognitive alignment

Investigating how causal language models respond to physical common-sense violations, and how their predictive distributions line up with human brain responses.

• Minimal pairs

  Formulated minimal pair test sets (e.g. key vs. banana) to analyze GPT-2's sensitivity to physical common-sense violations.

• Surprisal analysis

  Extracted next-token output logits and computed Shannon surprisal (−log₂ p), showing semantic anomalies yield 3× higher surprisal (M ≈ 11.1 vs. 4.1).

• Cognitive parallel

  Drew computational parallels between LLM predictive distributions and human N400 ERP brain responses.

Deep CNN for multi-class agricultural image classification

A deep convolutional network built for large-scale agricultural imagery, paired with an automated preprocessing pipeline.

• Accuracy

  Engineered a deep CNN classifying 75,000+ agricultural images with 98.5% test accuracy, outperforming standard baseline models by 14%.

• Pipeline

  Implemented an automated preprocessing pipeline with OpenCV for data augmentation, geometric feature extraction, and image normalization.