• Built a minimal deep learning framework from scratch with tensor operations, reverse-mode automatic differentiation, and dynamic computation graphs
• Implemented optimized GPU kernels using CUDA for matrix multiplication, convolution, and gradient propagation — 20x speedup over CPU
• Integrated modular components for end-to-end training of MLPs and CNNs with performance profiling

🧩 Technologies: Python, C++, CUDA, NumPy

• Designed a row-based vectorized query execution engine supporting sequential scans, inserts, updates, deletes, and B+Tree index scans with dynamic key-size template dispatching
• Architected hash join with Grace partitioning, nested loop/index joins, aggregation, external merge sort, and window functions
• Implemented optimizer rules to convert sequential scans into index scans and nested loop joins into hash joins

🧩 Technologies: C++, Database Internals

• Implemented a concurrent B+Tree index with thread-safe search, insertion, and deletion supporting node splitting, merging, and key redistribution
• Designed a tombstone buffer on leaf pages for deferred deletions as a simplified Bε-tree optimization
• Applied optimistic latch crabbing with read/write page guards for high-throughput parallel operations without a global latch

🧩 Technologies: C++, Concurrency, Data Structures

• Developed the Raft replicated state machine protocol with leader election, log replication, and commitment across concurrent peers via RPC
• Engineered AppendEntries and RequestVote RPCs with election timeouts, heartbeats, and election restriction for consistent leader election under network partitions
• Handled log consistency with conflict detection, nextIndex backtracking, and majority-based commit advancement

🧩 Technologies: Go, Distributed Systems, Concurrency

• Trained a Q-Former adapter bridging a frozen GPT-2 and a frozen Diffusion Transformer (DiT) for text-conditioned image generation on CIFAR-10 with classifier-free guidance
• Constructed learnable query embeddings with multi-layer self-attention, cross-attention over GPT-2 hidden states, and per-layer projection to DiT conditioning vectors

🧩 Technologies: PyTorch, Diffusion Models, Generative AI

• Designed the Live Sequence Protocol (LSP) for reliable UDP communication with sliding window, exponential backoff, and checksum-based integrity
• Built a client-server architecture with load balancing, fault tolerance, and task redistribution
• Achieved concurrency using goroutines and channels for scalable, efficient shared computing

🧩 Technologies: Go, Networking, UDP, Distributed Systems

• Built a real-time kernel with context switching, task scheduling, and mutex synchronization on ARM Cortex-M
• Implemented rate-monotonic scheduling (RMS) with Utilization Bound schedulability checks
• Designed thread control blocks (TCBs) for priority-based multi-threading management

🧩 Technologies: C, ARM Cortex-M, Embedded Systems, RTOS Design

• Implemented FFT from first principles using the recursive Cooley-Tukey algorithm, reducing complexity from O(N²) to O(N log N)
• Validated correctness against NumPy's backend and benchmarked performance vs. naive DFT to demonstrate scalability

🧩 Technologies: Python, CUDA, Signal Processing