AI Systems & Machine Learning Engineer
Specialized in parameter-efficient model alignment, high-throughput LLM serving infrastructure, generative computer vision, high-dimensional vector search, and real-time distributed spatial backends.

Languages & Core Frameworks:

Generative AI, PEFT & Retrieval:

Systems, Databases & Infrastructure:

1. vlm-spatial-eval — Spatial Robustness Benchmark & LoRA Alignment for VLMs

   • Domain: Multimodal Alignment, Vision-Language Robustness & Parameter-Efficient Fine-Tuning (PEFT)
   • Capabilities & Metrics: Diagnosed architectural degradation modes across CLIP, SigLIP, and SmolVLM under photometric noise. Engineered a conversational synthetic instruction dataset and trained low-rank adapters (q_proj, v_proj, $r=8$, 0.29% trainable params) with prompt loss-masking. Achieved an +8.3% absolute accuracy gain on balanced relational discrimination without catastrophic forgetting on open-domain visual tasks. Automated with 12 unit tests and GitHub Actions CI.

2. PagedInfer — High-Throughput Paged KV-Cache LLM Serving Engine

   • Domain: ML Systems Engineering, Memory Virtualization & Autoregressive Inference
   • Capabilities & Metrics: Implemented non-contiguous virtual memory block allocation for LLM key-value caches, eliminating memory fragmentation. Features continuous dynamic request batching, prefill/decode phase separation, and asynchronous streaming generation for high-concurrency client workloads.

3. Dual-Attention-GAN-LLIE — Generative Low-Light Image Restoration

   • Domain: Deep Learning Vision Research, Generative Adversarial Networks & Restoration
   • Capabilities & Metrics: Architected a custom U-Net generator featuring interleaved spatial and channel-wise attention mechanisms paired with a relativistic discriminator. Evaluated on the LoLI-Street benchmark dataset with rigorous quantitative validation across PSNR, SSIM, and perceptual LPIPS metrics.

4. VectorCore — High-Performance Vector Index & ANN Search Engine

   • Domain: Information Retrieval, High-Dimensional Vector Geometry & Embeddings
   • Capabilities & Metrics: Built a lightweight Approximate Nearest Neighbor (ANN) indexing system implementing optimized Cosine, Euclidean, and Dot-Product distance metrics. Delivers sub-2ms query latency with deterministic Recall@K across high-dimensional multimodal embedding distributions.

5. MCP-Mesh — Distributed Agent Gateway & Model Context Protocol Mesh

   • Domain: Agentic Systems, Tool Orchestration & Microservice Interoperability
   • Capabilities & Metrics: Centralized agent-to-tool integration mesh built on the Model Context Protocol (MCP). Implements dynamic tool discovery, bidirectional JSON-RPC schema validation, process isolation, and fault-tolerant agent execution graphs.

6. CTG_Shield_Spatial_Engine — Geospatial Risk Telemetry & Emergency SOS Dispatch

   • Domain: Geospatial Intelligence, Spatial Databases & Real-Time Telemetry
   • Capabilities & Metrics: Engineered an asynchronous FastAPI spatial telemetry engine backed by PostGIS spatial indexing (ST_Contains, ST_DWithin) for real-time hazard polygon intersections. Delivers sub-second emergency WebSocket broadcasts synchronized with the cross-platform native Flutter client (CTG_Shield_Mobile).