Hengkai Zheng

I’m a software engineering intern turned AI/ML builder, combining rigorous ML training with production-minded data pipelines. I’ve focused on turning messy, high-volume inputs into structured systems that support real-time search, analytics, and downstream use.

At Attencity, I worked as a Sole Developer to build web scrapers for Xiaohongshu, TikTok, and Douyin using Python Pandas, processing 3,000+ creator profiles/hour. I designed and maintained a PostgreSQL ingestion pipeline to store and query scraped creator data, enabling real-time search and bulk export.

In my projects, I’ve built hands-on AI systems with measurable evaluation: an Interactive Character Companion Agent using a 5-step stateful dialogue pipeline on Qwen2.5-1.5B-Instruct, with a two-tier memory system and FAISS semantic retrieval. I also curated a 4,793-sample SFT dataset across emotion categories and policy labels to train state-conditioned tone variation.

I’ve also implemented deep learning components from scratch to understand the “why” behind performance—building ResNet-34 + ArcFace metric learning, reporting EER, ROC-AUC, and TAR@FPR (achieving 2.55% EER on a held-out private leaderboard). Alongside that, I built a NumPy-based “MyTorch” framework with forward/backward passes, autograd, attention, RNN/GRU/CTC, and optimizers (SGD/Adam/AdamW).