Naicheng Deng

Founding AI Engineer at Sumi (Startup)

Founding AI Engineer building an end-to-end AI platform. Implemented a zero-shot PDF knowledge-graph pipeline with pgvector-based deduplication and incremental updates, achieving ~10 pages/s throughput; developed a stateful LangGraph runtime enabling reliable session recovery; implemented streaming hooks to reduce token usage by 40% vs ReAct while maintaining sub-3s P99 TTFT. Led backend & platform work including a modular-monolith backend on Cloud Run with CI/CD, delivered 30+ RESTful FastAPI endpoints, and established async testing/observability with pytest-asyncio, LLM mocking, and tracing, achieving ~80% test coverage.

Machine Learning Engineer at AI-Powered Adaptive Learning Platform

Engineered a Learning-to-Rank (LTR) system using transfer learning; pre-trained LightGBM on a proxydataset of 2.5M+ logs with rich behavioral signals and adapted the features space to the target environment. Deployed the model as a microservice using Async FastAPI and Docker; integrated Prometheus and Grafana for real-time system health and latency monitoring (<50ms). Architected an RAG framework combining Gemini embeddings with semantic deduplication to reduce LLM context window usage by ~40% while improving knowledge consistency. Built an end-to-end multimodal ETL pipeline to convert unstructured PDFs to knowledge graphs; optimized graph persistence latency (3.2s → 180ms) by replacing iterative N+1 inserts with atomic batch writes in PostgreSQL with ON CONFLICT handling. Implemented GitOps workflows using MLflow for model registry and GitHub Actions for CI/CD; established caching strategies on Google Cloud Run to accelerate build-and-deploy cycles (15min → 3min).

Machine Learning Engineer Intern at Qishu Data (Shanghai) Co., Ltd.

Developed a daily ETL pipeline integrating PostgreSQL telemetry with HubSpot CRM data, and engineered behavioral features to identify high-intent trial users. Built and deployed a LightGBM lead-scoring model (Precision@K) on imbalanced data, delivering ~16% uplift over heuristic scoring in daily outreach prioritization.

Research Engineer (High-Performance Computing) at University of Waterloo

Implemented a custom PyTorch C++/CUDA operator for sparse interaction matrices, delivering ~3x end-to-end training speedup on AWS NVIDIA T4 GPUs. Built an RLE-based sparse storage and checkpoint pipeline with chunked AWS S3 transfers, reducing memory footprint to ~10% of CSR for matrices up to 1e6 × 1e6 interactions. Validated numerical correctness with gradcheck and finite-difference tests; instrumented CloudWatch loss/gradient/throughput metrics to maintain stability of long-running training.

Researcher (High-Performance Computing) at University of Waterloo

Optimized particle neighbor search from O(n^2) to O(n) using spatial hash grids; reduced memory usage from 200GB to 15GB, enabling large-scale simulations on cloud GPUs. Developed a neural-network solver for functional optimization problems in physics simulations; implemented custom CUDA kernels via CuPy for gradient computation and integrated with PyTorch Autograd for nested differentiation, achieving ~15× speedup (3 days → 6 hours). Implemented fault-tolerant checkpointing with HDF5 to recover long-running jobs in ephemeral containerized environments.

Research Engineer (Machine Learning) at University of Waterloo

Developed a physics-constrained sequence generator with LSTM and WGAN-GP, using gradient-penalty stabilization to mitigate mode collapse. Built a reproducible experimentation workflow with MLflow tracking, NumPy/SciPy metric post-processing, and CLI-based checkpoint/resume support.