Projects

Projects

Reproducible, measured work — each repo states what it is, what it is not, credits the upstream tools it drives, and ships measured results rather than claims. GPU work runs on 4× H100 (NVLink) and Blackwell workstation hardware.

Tags: CUDA Serving Benchmarks Agents RAG Privacy

LocalMind-Core-GX10 — enterprise multi-agent AI platform

Sole developer · SYNCROBOTIC · Sep 2025 – Jun 2026 · Python / TypeScript / Rust / vLLM / LightRAG / NestJS / DGX

Polyglot architecture — a Rust performance core for dispatch and caching, a NestJS + TypeScript API gateway, and Python ML services — with a unified provider router and an agent layer over LightRAG hybrid retrieval (vector + keyword + knowledge graph). One platform replaces ad-hoc per-team integrations, with built-in tracing, regression testing, and safety validation. Shipped at two enterprise customers.

AgentsRAGServing

GPU inference benchmarks & kernel studies

CUDA / Triton / PyTorch / Nsight · in progress — code release planned

LLM serving internals, hands-on: KV-cache behaviour, quantization trade-offs, Flash Attention; hand-written kernels (tiled GEMM, WMMA Tensor Core matmul) profiled with Nsight Compute to connect kernel-level decisions to end-to-end serving latency.

CUDAServingBenchmarks

TensorRT-LLM + Triton multi-GPU serving

TensorRT-LLM / Triton / vLLM / 4× H100 NVLink · release planned once results land · notes

Reproducible build → serve → benchmark harness: TensorRT-LLM engines on Triton (TP=4), measured head-to-head against vLLM under matched concurrency.

ServingBenchmarks

NCCL collectives benchmark

NCCL / nccl-tests / CUDA / 4× H100 NVSwitch · github · notes

Bus-bandwidth micro-benchmarks for all-reduce / all-gather / reduce-scatter on 4× H100 NVLink, analysed against the theoretical link budget. Measured: all-reduce 366 GB/s (77% of NVLink); NVLS > Ring > Tree; protocol study (Simple/LL128/LL).

CUDABenchmarks

NIM agent blueprint

NVIDIA NIM / RAG / agents / OpenTelemetry / FastAPI · github · notes

Agentic RAG reference architecture on NVIDIA NIM microservices (LLM + embedding + reranker) with a plan → retrieve → generate → validate loop and a built-in eval harness. Measured: retrieval recall@3 100%, and 0% hallucination on out-of-corpus questions with a guarded prompt vs ~40% without it (ablation).

AgentsRAG

Blackwell Tensor Core kernels

CUDA / Tensor Cores / WMMA / Nsight / H100 + Blackwell workstation · release planned once results land · notes

Hand-written CUDA GEMM kernels (naive → tiled → WMMA Tensor Core), benchmarked across Hopper (sm_90) and Blackwell (sm_120) as a fraction of the cuBLAS ceiling.

CUDABenchmarks

Federated learning lab

Python / PyTorch / FL / DP · github · notes

From-scratch federated learning — FedAvg / FedProx / SCAFFOLD, DP-SGD and secure aggregation, plus FedPer / Byzantine-robust / FedAdam / FedLoRA. 33/33 tests, literature-cross-validated, with honest negative results on Non-IID MNIST.

PrivacyFederated Learning