Tadiwa Mandava
Education
University of Oklahoma
B.S. Computer Engineering, Minor in MathematicsNorman, OK
Experience
Software Engineering Intern
K20 Center, University of OklahomaNorman, OK
- Develop and deploy front-end and back-end web applications and Unity/C# educational games used by K–12 teachers and students; test across Center projects, reproducing and filing defects
- Built and shipped Hive, a classroom platform for curating and publishing boards of links, images, and video; owned the stack from PostgreSQL schema through REST API to deployed React SPA
IT Service Delivery Specialist
University of Oklahoma ITNorman, OK
- Administer identity and access management and Microsoft 365 services across 30,000+ user accounts, automating account provisioning and recurring administrative workflows with PowerShell
- Operate on a five-person cross-functional team handling ticket-driven telecom, account, and backend operations requests, sustaining a 98.4% satisfaction rate
Projects
- Built a crash-safe, concurrent key-value storage engine with an LSM-tree architecture: write-ahead log, immutable SSTables, Bloom filters, and background compaction, exposing Redis-compatible commands
- Validated compaction, WAL recovery, and concurrency correctness with property-based tests against a reference HashMap
- Implemented GPT-2 inference from scratch in C++, including tokenization, transformer blocks, multi-head attention, KV caching, sampling, and tensor ops, without an ML inference framework
- Optimized CPU inference with cache-aware matmul, SIMD, and multithreading; implemented INT8/INT4 quantization, validated against a Hugging Face reference
- Implemented an SGEMM optimization ladder (coalesced loads, shared-memory tiling, register blocking, thread coarsening, vectorized loads), validating each kernel against cuBLAS on non-power-of-two shapes before benchmarking
- Profiled each stage in Nsight Compute using achieved occupancy, DRAM and SM throughput, warp stall reasons, and register pressure, selecting each optimization from the binding hardware constraint rather than intuition
- Implemented a fused FlashAttention kernel that tiles Q/K/V through the GPU memory hierarchy instead of materializing the full attention matrix, exposed as a PyTorch custom operator verified against SDPA
- Built and shipped Hive, a classroom platform for curating and publishing boards of links, images, and video; owned the stack from PostgreSQL schema through REST API to deployed React SPA
Skills
- Languages
- PythonARM assemblyCC++CUDA C++JavaScriptRustSQLSystemVerilogTypeScript
- Frameworks & tools
- AdonisJSarm-none-eabi-gccAVX2Basys 3criterionEmscriptenGDBGTKWaveNsight ComputeOpenMPOpenOCDPostgreSQLproptestPyTorchReactSTM32F446tokioVerilatorViteVivado