I'm an AI Infrastructure Engineer who builds fast, cost-efficient LLM serving systems. I take inference research — KV-cache disaggregation, FP4 quantization, and RDMA-scale compute — and turn it into measurable production wins on A800/H200 clusters. I'm an M.Eng. candidate at NWPU and the author of NVFP4-DiT (IEEE TMM, under review).
- 🛠️ Building a 10-project open-source AI-infra portfolio to sharpen and showcase production engineering.
- 📄 Writing my M.Eng. thesis on multimodal deepfake detection.
- 💡 Open to AI Infrastructure / LLM-serving engineering roles.
flowchart LR
BD[Bangladesh] --> CN[China: Xi'an]
CN --> BS[NWPU: BSc in CST]
BS --> MS[NWPU: MEng in SWE]
MS --> AI[AI Infrastructure Engineer]
MS --> RS[NVFP4-DiT Research]
AI --> PF[10-Project Open-Source Portfolio]
RS --> PF
PF --> FUT[The Road Ahead]
- AI Infrastructure — KV-cache architecture, prefill/decode disaggregation, SGLang/vLLM, RDMA, A800/H200 clusters.
- Low-Precision Inference — FP8/FP4 quantization for LLMs and diffusion transformers.
- M.Eng. Software Engineering @ NWPU — LLM inference optimization & multimodal AI.
- LLM Serving at Scale — KV-cache disaggregation (PD split), prefix caching & continuous batching on SGLang/vLLM across A800/H200.
- Low-Precision Inference — FP8/FP4 quantization for LLMs and diffusion transformers; author of NVFP4-DiT (IEEE TMM, under review).
- Distributed GPU Systems — RDMA-based multi-node training/inference and cluster orchestration.
- GPU Kernels — CUDA / Triton / FlashAttention kernels for attention and GEMM.
- Multimodal AI — audio-visual-temporal fusion for deepfake detection and video generation.
| Metric | Result | Context |
|---|---|---|
| KV-cache hit rate | 92.27% | prefix-cache tuned serving |
| TTFT reduction | 8.3× | 4× A800 PD-disaggregation |
| Memory reduction | 4× | FP4 quantization (NVFP4-DiT) |
| Thesis | 70 pages | multimodal deepfake detection |
| Scholarship | Chinese Gov. Scholarship | NWPU M.Eng. |
AI Infrastructure Engineer Intern · InfiX.ai · Shenzhen, China · Apr 2026 – Jun 2026
| Degree | University | Period |
|---|---|---|
| M.Eng. Software Engineering | Northwestern Polytechnical University (985/211) | Sep 2024 – Mar 2027 |
| B.Eng. Computer Science & Technology | Northwestern Polytechnical University (985/211) | Sep 2020 – Jul 2024 |
- AI Infrastructure & LLM Inference Systems (KV-cache, PD disaggregation, RDMA)
- Multimodal AI (visual + audio + temporal fusion, deepfake detection, video generation)
- Low-Precision Quantization (FP8/FP4 for diffusion transformers and LLMs)
- GPU Kernel Optimization (CUDA, Triton, FlashAttention, GEMM)
- LLM serving infrastructure and distributed systems
- Multimodal AI and computer vision research
- Open-source AI tooling and frameworks
- Always happy to connect via email or LinkedIn
A curated set of 10 production-quality, fully-tested, Dockerized AI-infra projects — each with a clear architecture, quickstart, and CI. Showcased in ai-infra-portfolio.
| Project | Area | What it does |
|---|---|---|
| llm-gateway | LLM Gateway | OpenAI-compatible proxy: per-key rate limits, cost caps, caching, provider fallback. |
| mini-serve | Model Serving | Inference server: dynamic batching, SSE streaming, Prometheus metrics. |
| gpu-exporter | Observability | NVIDIA GPU Prometheus exporter + Grafana dashboard. |
| train-launcher | Training Ops | Fault-tolerant multi-node PyTorch launcher with auto-retry. |
| model-registry | MLOps | Content-addressed model/dataset registry (mini-MLflow) with aliases. |
| kv-cache-sim | Inference Eff. | KV-cache eviction + prefill/decode disaggregation simulator (TTFT/TPOT). |
| llm-bench | Benchmarking | LLM benchmark/eval harness: TTFT, TPOT, throughput, cost model. |
| quant-playground | Low Precision | FP8/FP4/INT8/INT4 bit-packing + numpy MLP quantization demo. |
| llmoops-trace | Observability | OpenTelemetry LLM tracing collector + Grafana dashboard. |
| rag-pipeline | Retrieval (RAG) | Offline RAG toolkit: ingest → chunk → embed → ANN search → rerank. |
| Project | Description |
|---|---|
| NVFP4-DiT | 4-bit low-precision audio-guided video diffusion transformer (IEEE TMM, under review). |
| DeepfakeAudioVisualTemporalDetector | Multimodal deepfake detection: EfficientNet + FFT + MFCC + Transformer + Attention Fusion. |
| c-compiler-frontend | 4-stage C compiler front-end: lexical, syntax, semantic analysis + three-address code generation (Flex/Bison). |
| Online-Portfolio | Personal portfolio website. |
- NVFP4-DiT: 4-bit Audio-Guided Video Diffusion Transformers — IEEE TMM, under review.
“Inference is where the model meets the metal — I make that intersection fast, cheap, and observable.”