Building high-performance foundations for Next-Gen AI. Expert in squeezing maximum TFLOPS out of consumer hardware and scaling LLM inference across distributed clusters.

Currently based in Curitiba, Brazil (GMT-3). Open to relocation worldwide.

Full C++/CUDA Runtime

• Transformer inference engine in pure C++/CUDA, no PyTorch execution path.
• 32.4k tok/s peak, achieving 87% of INT8 Tensor Core ceiling on RTX 2070.
• Custom Flash Attention, cuBLASLt INT8, CUDA Graphs, Megatron-style TP.
• Runs TinyStories 110M and TinyLlama 1.1B end-to-end.

PTX-Level GEMM Engineering

• Manual PTX instruction emission for Tensor Core GEMM.
• Beats cuBLAS at 2048×2048×2048; achieves 60-65% at 8192×8192×8192.
• Auto-tuning over WARPS / STAGE_K / ILP_DEPTH configuration space.

CUDA Kernel + End-to-End LLM Integration

• Bit-perfect INT4 weight-only quantization kernel using raw PTX assembly.
• Achieved 1.000 Cosine Similarity against FP16 baselines.
• End-to-end Qwen2-0.5B integration: -14.7% VRAM with coherent output.
• Documented SM75 structural limits (no cp.async, no INT4 MMA operands).

Compiler MVP with CUDA Codegen

• Full pipeline: IR → CUDA codegen → nvcc → GPU execution.
• 9 Transformer fusion patterns (GEMM+Bias+GeLU, LayerNorm fusion, etc.).
• 10.5 TFLOPS peak on RTX 2070 (37% of SM75 theoretical peak).
• Templated CUDA generation via Jinja2 with type/shape specialization.

LLM Scaling & Distributed Systems

• Multi-process runtime implementing Megatron-style Tensor Parallelism and Pipeline Parallelism.
• GPipe Scheduler: Fill-drain micro-batching to minimize idle bubbles.
• KV Cache Sharding: Distributed attention state to scale beyond single-GPU VRAM limits.
• Bit-identical validation on Qwen2-0.5B weights.

Production Serving Runtime

• OpenAI-compatible inference server (/v1/completions, /v1/chat/completions).
• Continuous batching, SSE streaming, KV cache management.
• Prometheus metrics, graceful shutdown, health checks.
• 33/33 unit tests + 6/6 end-to-end scenarios passing.

LLM-as-a-Service Platform (Production-Ready)

• High-concurrency backend with OpenAI-compatible API.
• ACID Billing: PostgreSQL-backed financial ledger with Decimal precision.
• Distributed Rate Limiting: Redis ZSET sliding window (multi-worker safe).
• Tool calling execution layer, Stripe integration, admin dashboard.
• Production stack: Docker Compose, Gunicorn, Prometheus.

Mathematical Model Compression

• From-scratch implementation of AWQ and GPTQ algorithms.
• Validated intelligence preservation on Qwen2-0.5B with 26.05 Perplexity at 4-bit.
• Deep-dive study on Hessian-based error compensation vs activation salience.

• flash-attention-sm75-register-fused: Register-fused FlashAttention for Turing (5.0 TFLOPS peak, roofline-validated).
• flash-attention-sm75: FlashAttention v1 forward pass in CUDA for SM75.
• fused-backward-kernels-sm75: Fused GEMM+Bias+GeLU backward kernels (3.1× faster than autograd at M=1024).
• rag-inference-stack: High-speed local RAG using ChromaDB and sentence-transformers.
• speculative-decoding-runtime: Accelerated inference (1.41× speedup) via draft-model verification.
• vlm-inference-runtime: Multi-modal deployment wrapper for Vision-Language Models.
• lora-inference-runtime: Multi-tenant serving with sub-3ms adapter hot-swapping.

AI Infrastructure Engineer (Contract) — Confidential Leading Chinese LLM Company (Under NDA) | Sep 2025 – Mar 2026
Built Rust-based ingestion pipelines with Tokio async and backpressure-safe systems supporting an in-production LLM stack.

Software Architecture Lead & Operations Manager — ZBrasil (Cross-border E-commerce) | Apr 2025 – Aug 2025
Led backend architecture and GPU infrastructure decisions.

• Contributed to LLM Inference Handbook (llm-inference-at-scale) — module on observability signal coverage for LLM serving.

• LinkedIn: linkedin.com/in/joaofelipescheuer
• Email: johnfelipe13@gmail.com
• WhatsApp: +55 41 99705-4864
• WeChat: JohnScheuer7
• Location: Curitiba, Brazil (GMT-3) — Open to Relocation Worldwide

Code the math, scale the intelligence.