ML/MLOps engineer focused on production-grade systems at the intersection of LLM security & agentic infrastructure, computer vision, and geospatial ML — with 12 years building deep learning systems, scalable data pipelines, software backend systems, and cloud infrastructure (GCP/AWS). That work spans 30 public repositories, including 5 production-shaped platforms with real CI/CD, Terraform-provisioned infrastructure, and tagged releases rather than perpetual drafts.

I build the full stack an organization needs to run AI systems in production. Data foundations, LLM serving and GPU orchestration, computer-vision pipelines, and observability with incident response — end to end. Where it matters, the numbers are measured, not estimated: one lakehouse pipeline processed 42,972 records at a 91.8% acceptance rate with zero post-gate rejections; another platform runs 45 Docker Compose services behind 41 automated test files and 6 active CI workflows; a third ships 17 services behind a single multi-warehouse activation contract. Success is serving reliability, GPU cost-safety, and pipeline acceptance rates — not notebook demos.

Production-shaped multi-backend LLM serving and GPU-orchestration platform. One unified serving contract across vLLM, Triton Inference Server, Ray Serve, KServe, and BentoML — with real GPU infra patterns (MIG partitioning, Kueue multi-tenant scheduling, Karpenter autoscaling) validated in CI so no GPUs are burned in the pipeline. Tagged v1.2.0 across 22 phases: SageMaker/Bedrock integration, LoRA/PEFT fine-tuning, DVC data versioning, pluggable MLflow/W&B tracking, TensorRT-LLM templates, GPTQ/AWQ/FP8 quantization, cost-per-token tracking, and a LangGraph advisor agent that is tool-grounded and non-fabricating — CI asserts every number it states is backed by Prometheus/benchmark evidence from that run.

LLM Serving · GPU Orchestration · Kueue · Karpenter · KServe · vLLM · Triton · Ray Serve · Quantization · MLOps

Vulcan's sibling platform: a production-shaped fleet telemetry, data-quality, MLOps, and observability system. Tagged v1.0.0, with its own CHANGELOG calling it the "first production-shaped release" across Phases 0–15 — a 45-service Docker Compose stack, 41 test files (including Kafka integration tests), and 6 active CI workflows. Devices stream into a Kafka/Redpanda bus; Ray and Flink harden the data path; Iceberg + Trino form the lakehouse; Dagster/MLflow orchestrate; drift-monitor, an OPA-backed incident-engine, and OpenTelemetry close the operational loop; a read-only, Qdrant-RAG AI copilot ships with its own eval harness. Polyglot monorepo (Go for incident-engine / API gateway / CLI, Python for ingestion / drift / copilot, Next.js dashboard, Terraform / Helm / ArgoCD for infra).

Fleet Telemetry · Iceberg · Trino · Dagster · Kafka · Ray · Flink · OPA · OpenTelemetry

Multi-warehouse fleet-intelligence platform. Tagged v1.2.0, with all 20 phases (0–19) complete — including a golden-path chaos end-to-end test run against the live Compose stack. Camera/sensor ingest splits into a PySpark medallion lakehouse (bronze → silver → gold, dbt-modeled) and an OpenCV/ONNX YOLO-family CV service that routes low-confidence findings to a Django control-plane review queue. Gold data fans out through one activation contract to Redshift and Snowflake, mirrors to Databricks/Azure ADLS for disaster recovery, and feeds a Vue 3 + Three.js digital-twin cockpit; a per-asset circuit-breaker incident-engine runs OPA/Rego trip policies; a tool-grounded AI copilot answers only from evidence it can cite. 17-service Docker Compose stack, 19 test files, 2 CI workflows.

PySpark · Databricks · dbt · OpenCV · ONNX · Django · Vue · Snowflake · Redshift

Offline AV perception & auto-labeling platform. Tagged v0.2.0, with all 11 phases complete. An 8-stage forge CLI pipeline runs end to end locally: ingest (nuScenes → Parquet lake) → detect2d (Faster R-CNN) and detect3d (PointNet-style) → track (SORT: Kalman + Hungarian IoU) → fuse (calibrated sensor fusion) → label (active learning / pseudo-labeling) → evaluate (BEV distance, mAP against held-out ground truth) → curate (LanceDB dedup) → visualize (rerun.io / Foxglove MCAP). A parallel cloud path (S3 → Lambda → SQS → DynamoDB → EventBridge → Step Functions → ECS Fargate → Glue → Athena, 11 tables) is Terraform-defined and CI-verified — deliberately never applied against live AWS, the same cost-safety policy as Vulcan, PRISM, and hydra-data-factory. Detection heads are documented honestly as randomly-initialized research baselines, not tuned models.

Computer Vision · PyTorch Lightning · Sensor Fusion · Ray · LanceDB · Terraform

AI-native defense-in-depth gateway for LLM apps and agents. Tagged v0.3.1, with all 12 build-order stages complete — an 11-service Docker Compose stack, 53 test files, 3 CI workflows (ci, release, security). A Go gateway routes requests through Python input-defense detectors, a Go + CEL policy engine, and a provider-agnostic model router, then output-defense detectors (plus an LLM judge) before any response releases; a separate agent-gate enforces tool/MCP permissions with taint tracking; every decision is Ed25519-signed into a Postgres audit trail. Security decisions fail closed — an outage returns 502, never an unchecked response — documented explicitly in FAILURE_MODES.md. It also publishes its own adaptive red-team result rather than a vendor catch-rate demo: real-model hardening cuts round-1 bypass rate (10.8% → 9.2%), but overall bypass rate under sustained adaptive attack stays flat at ~48% — reported honestly rather than smoothed over.

LLM Security · Policy-as-Code · CEL · Prompt Injection · Tamper-Evident Audit · Red-Teaming · Go · Python

Production-validated AV telemetry lakehouse — the data/lakehouse foundation Argus grew out of. Mock fleet telemetry simulator, two-layer Pydantic + Pandera contract validation with dead-letter-queue isolation, Terraform-provisioned S3 + Glue + IAM, and PyArrow/Parquet columnar storage with Hive partitioning, deployed via Docker Compose. Verified production-run metrics: 42,972 records ingested, 91.8% acceptance rate, 0% post-gate contract rejections.

Data Contracts · Pandera · PyArrow · Parquet · Terraform · S3 · Glue · Docker Compose

• AegisFlow — Autonomous multi-agent incident response engine (Temporal, PydanticAI, OpenTelemetry) that triages observability alerts and plans mitigations.
• smart-assistant — Personal knowledge assistant and workflow agent.
• edgevision — Edge CV security/anomaly detection with YOLO + ByteTrack.

• terra-OBIA — Object-Based Image Analysis platform for forestry stand delineation, wetland classification, and land cover mapping at province scale.
• terra-obia-etl — ETL pipeline harmonizing GeoNB/GNB Open Data into labeled training polygons.

• NEXUS-CV — Real-time multi-modal computer vision platform with live observability and full MLOps lifecycle.
• multi-modal-od — Sensor-fusion object detection (radar, LiDAR, camera) for autonomous driving.
• sentinel-ray — Distributed camera telemetry ingestion with Ray, statistical drift detection, and automated incident orchestration.
• vanguard-telemetry-monitor — Fleet telemetry simulation and observability platform with production-like anomaly injection.
• vit-fastapi-cloud-deploy — FastAPI image classification service, evolved into a cloud-deployed Vision Transformer pipeline.
• cuda-optimization — Deep learning optimization guide adapted for Apple Metal Performance Shaders.

• task-master — Production FastAPI + PostgreSQL REST API with Docker and CI/CD.
• bert-inference — Containerized DistilBERT sentiment-analysis inference service.

Production-quality platforms are resource-efficient and evidence-grounded by design — not bolted on after the model works. I treat serving contracts, GPU cost-safety, and observability as first-class systems problems: unify backends, schedule capacity without burning money in CI, validate data at the gate, and close the loop with metrics that survive contact with production. Success is measured by deployment metrics — serving reliability, GPU utilization under multi-tenant constraints, pipeline acceptance rates, and operational signal fidelity — not theoretical benchmarks that never leave the lab.

📍 Fredericton, NB, Canada · 🌐 hamidmatiny.github.io