Work

AI reasoning systems for tasks of high cognitive complexity.

Founding infrastructure engineer. Built the platform infrastructure 0→1, the full inference/deployment/observability stack, before the team grew around it. The company launched on it into the AWS and Azure Marketplaces and the AWS Global Startup Program, and earned a place in Microsoft's invite-only Pegasus program.

• –Built SYMI's execution sandbox. Agents take real actions across email, CRM, and external systems, and you can't trust the model's calls. So every session runs walled off: Linux namespaces, cgroups, seccomp, microVM boundaries.
• –Put eight tenants on one GPU and held inference under 50ms doing it: MIG partitioning, speculative decoding, 8:1 sharing without the latency tax, on a vLLM stack running 12x throughput at 60% less GPU memory.
• –Did the GPU acceleration on the knowledge graph the platform retrieves from. Graph traversal and similarity ranking run on the GPU, so retrieval scales to millions of nodes without becoming the bottleneck inference waits on.
• –Designed the deployment model: one build, every target. SaaS multi-tenant, single-tenant inside a customer's own AWS, Azure, or Google Cloud account, or fully on-prem and air-gapped. Kubernetes-native with no managed-cloud service dependencies, so the same artifact ships to all three major clouds and a private datacenter with no rewrite.
• –Built the model serving behind Constellation, the verification layer between draft and final response: a DAG of agents parallelized where dependencies allow, bounded so cognition stays predictable. I also built the model-as-judge harness that grades every customized model against a gpt-oss-120b baseline before it ships.
• –Led the compliance effort to SOC-2 Type II and ISO 27001, GDPR and CCPA on top: single-tenant isolation, JWT and SSO auth, RBAC down to resource:action grants, and a governance kernel that stops the model executing anything it wasn't approved to. Cut infrastructure cost 70% and held 99.9% uptime running it.

Python, Kubernetes, PyTorch, Ray, vLLM, TensorRT, TensorRT-LLM, torch.compile, CUDA, Custom CUDA Kernels, TransformerEngine, FlashAttention, Nsight Compute, Nsight Systems, ArgoCD, Helm, Kustomize, Prometheus, OpenTelemetry, Grafana, AWS, Docker, GitOps, CI/CD, GPU Scheduling, Mixed Precision, ONNX Runtime

Rearchitected batch systems into real-time streaming. Built an observability platform from scratch and presented it to exec leadership. Optimized SageMaker endpoints until inference costs stopped being a line item anyone questioned.

Apache Airflow, Python, AWS, SageMaker, gRPC, Databricks, Kubernetes, Docker, Helm, Terraform, Prometheus, OpenTelemetry, Distributed Tracing, CI/CD Pipelines, MLflow, Model Serving, ETL Pipelines, SQL, Load Balancing, IAM

25TB of market data. Every day. Sub-millisecond latency. I built the tick-level processing system behind $2M+ in annual trading decisions. Cut order execution latency by 78%.

C++, Python, Apache Kafka, Apache Spark, Low-Latency Networking, GPU Profiling, TLS, DNS, Network Optimization, Real-Time Analytics, gRPC, Bash

Seven business units with zero interoperability. I wired ML platforms into legacy Teradata and Oracle systems. Cross-system compatibility up 65%. Data quality up 85%.

Python, SQL, Teradata, Oracle, ETL, Data Pipelines, Data Governance, Java

proprietary · 4 systems

Proprietary Work

Closed source. Built privately.

Inference serving for video world models, built on spatiotemporal PagedAttention: KV paging extended across the time axis so successive frames share cache the way tokens do within a sequence. A DACB scheduler over a disaggregated five-pool topology and a three-tier KV hierarchy hold decode latency under 50ms at 85% GPU utilization.

Go, CUDA C++, Python, PagedAttention, FlashAttention, Kubernetes, gRPC, Raft Consensus, OpenTelemetry, GPU Memory Management, Kernel Fusion, Occupancy Optimization, Model Serving Architecture, Quantization (FP16), Nsight Compute

A GPU inference hypervisor in Rust and hand-written PTX/SASS, with no Linux kernel, CUDA driver, or Python runtime in the path. Bypassing those layers cuts stack overhead 95% and holds stack latency to 15-70 microseconds. Cold-starts in single-digit milliseconds where a Linux and CUDA-driver stack takes tens of seconds.

Rust, Assembly, CUDA, Bare Metal, GPU Programming, Warp-Level Primitives, GPU Memory Management, Custom CUDA Kernels, Nsight Systems, Profiling

A GPU-native vector database that runs search on tensor cores. 3.5M vector queries per second per GPU, 1.8 microsecond p50 latency, 500K inserts per second, and roughly 10x lower cost per query than managed cloud vector providers.

CUDA, Tensor Cores, Rust, NVLink, GPUDirect, Lock-Free Algorithms, GPU FinOps, Kernel Fusion, Occupancy Optimization, Custom CUDA Kernels

Declarative database infrastructure. No migrations. Bidirectional state convergence with SMT-verified invariants. O(n log n) complexity guarantees. Parallel DDL via dependency graph. Adopted by an internal-tooling team at Amazon.

Rust, SMT Solver, PostgreSQL, Formal Verification, Graph Theory, CI/CD, Distributed Systems

open source · 6 repos

Open Source

Multi-agent LLM serving for Apple Silicon's unified memory. existing systems were designed for discrete GPUs where weights must be copied over PCIe. on M-series chips, CPU, GPU, and Neural Engine share one physical pool -- that copy is unnecessary. PHANTOM eliminates it. 10 agents sharing a 50-page document: prefix stored once, not 10 times. DualRadixTree copy-on-write KV cache. MESI coherence formally specified in TLA+. formally verified scheduler. M0 proven: zero-copy GPU pipeline working end to end.

Rust, Apple Silicon, Metal, Unified Memory, TLA+, Formal Verification, Multi-Agent Systems, KV Cache, Copy-on-Write, Neural Engine

Autonomous GPU cluster orchestration. Replaces on-call SRE judgment with a hierarchy of specialized agents that perceive hardware degradation before it becomes failure. Topology-aware scheduling, not just GPU counts. Heals running training jobs without restart using NCCL 2.27 Communicator Shrink. Validated against the Alibaba Cluster Trace dataset. Every benchmark reproducible from a single command.

Rust, Python, NCCL, Kubernetes, GPU Scheduling, Distributed Systems, Multi-Agent Systems, Fault Tolerance

Task-Aware Sparse Fine-Tuning. Co-trains LoRA adapters with block-sparse attention gates. 2-5x decode throughput at 70-85% sparsity. 676 tests passing. Cuts inference costs without pretending accuracy doesn't matter.

Python, PyTorch, LoRA/QLoRA, CUDA, FlashAttention-2, Block-Sparse Attention, vLLM, Quantization, Model Compilation, Transformer Architecture Optimization, Mixed Precision, Gradient Checkpointing

Kubernetes scheduler plugin. Network topology-aware, cost-based, and performance-driven pod placement. The scheduler your cluster should have shipped with.

Go, Kubernetes, Docker, Helm, GPU Scheduling, Cold-Start Optimization, Multi-Region, Ingress, Load Balancing

Generate production-ready Airflow DAGs from natural language. 45 tokens/sec on CodeLlama 7B. ~700ms on an M2 Pro. No API calls. No cloud dependency. Your DAGs, your machine.

Python, Apache Airflow, LLMs, Ollama, vLLM, Model Serving

Neural network training in the browser. WebGPU compute shaders. No server. No Python. The model trains on your GPU, in your tab.

TypeScript, WebGPU, WGSL