AI Engineer | Computer Vision & Multimodal | Systems Inference & MLOps
AI Engineer with 5 years of experience building and delivering production computer vision pipelines for geospatial and municipal intelligence.
My work bridges high-resolution visual modeling with generative inference—designing hierarchical vision pipelines while engineering resource-constrained serving platforms governed by strict reliability contracts.
| Domain | Core Stack & Frameworks |
|---|---|
| Vision & Multimodal | YOLOv8 / YOLOv11, SAHI, OpenCV, PyTorch, Vision-Language Models (VLMs) |
| Inference & Systems | vLLM, KServe, Envoy AI Gateway, AWQ INT4, FP8 KV-Cache, k3s, Docker |
| MLOps & Data Engine | Labelbox SDK, DeepEval, GCP Vertex AI, Kubeflow Pipelines (KFP v2), Prometheus, Grafana |
| Backend & Contracts | Python 3.12, FastAPI, LangGraph, AsyncIO, Pydantic v2, AST Parsing (sqlglot), SQL |
Kubernetes-native serving architecture engineered for low-latency, deterministic LLM inference under compute constraints.
- Architecture: Single-node k3s cluster deploying KServe (
LLMInferenceService), vLLM runtime, and Envoy AI Gateway with OpenAI-compatible API contracts. - Optimization: Calibrated compute budgets using AWQ INT4 quantization, FP8 KV caching, prefix caching, and token-aware admission control to prevent OOM bottlenecks.
- Telemetry & Delivery: Instrumented real-time Time-To-First-Token (TTFT), inter-token latency, and KV-cache saturation via Prometheus/Grafana; deployed via hermetic GitOps pipelines.
Deterministic reliability bridge enforcing AST-based query validation, read-only safety, and schema inspection between LLMs and SQL engines.
- AST Validation: Traverses Abstract Syntax Trees (
sqlglot) to block destructive commands (DROP, DELETE, UPDATE, ALTER) and malformed syntax ahead of database execution. - Self-Correcting Loops: Implements schema-aware reflection routines that guide local and cloud LLMs to resolve schema/column mismatches autonomously.
- Predictable Contracts: Backed by Pydantic v2 schema enforcement and async state machines for predictable execution paths.
End-to-end production modeling and dataset engineering for large-scale aerial and satellite inspection.
- Backbones & SAHI: Fine-tuning and benchmarking detector backbones (YOLOv8/v11, OpenCV) integrated with Slicing Aided Hyper Inference (SAHI) to resolve small-object recall bottlenecks across multi-gigapixel scenes.
- Multimodal Compliance: Designing two-stage architectures that extract localized visual patches via detector backbones to feed downstream Vision-Language Model (VLM) evaluation pipelines.
- Data Operations: Programmatic dataset conversion tooling, structured ontologies, and consensus auditing on Labelbox across 138K+ high-resolution scenes.
- GitHub Copilot Individual Certification (GH-300) — Microsoft / GitHub (Active: 2025–2027)
- Microsoft Certified: Azure AI Fundamentals (AI-900) — Microsoft (Active: 2025)
- AWS Certified Machine Learning – Specialty — Amazon Web Services (Alumni: 2021)
- Technical Community Lead & Advisory Lead @ DevOps Visions: Architecting proof-of-concept demonstrations focusing on cloud infrastructure patterns, automated data pipelines, and system observability.
- Engineering Mentorship (Elmentor Program): Mentoring peer engineers on production AI roadmaps, cloud architecture patterns, and transitioning into production engineering.
- Technical Speaker: Facilitated hands-on workshops and sessions on AI-assisted developer workflows (GitHub Copilot across the SDLC) and engineering metrics.
- LinkedIn: linkedin.com/in/ayman-aboghonim
- GitHub: github.com/aymanaboghonim
- Email: aymanaboghonim@gmail.com