Machine Learning Engineer - Defendec/Reconeyez

Job Description
The Role
We're looking for a Machine Learning Engineer to own and evolve our models and ML infrastructure behind our actor-detection and visual-verification pipeline. This is the team that decides what our cameras "see" — from the object-detection models that flag intrusions, to the duplicate-suppression logic that stops a parked car from firing alarms all night, to the next generation of vision-language models we're bringing in for richer scene understanding (fly-tipping detection, license plates, image-quality scoring).
This is a hands-on engineering role, not a research-only one. You'll train and optimize models and get them running reliably in production — building the data pipelines(and MLOps), serving infrastructure, and evaluation harnesses that turn a notebook experiment into something that survives contact with real field imagery (day/night, IR/RGB, weather, bad signal). You'll also help shape where we take agentic and LLM/VLM capabilities next.

What You'll Do
• Train, fine-tune, and evaluate computer-vision models (object detection, image quality, static-object/duplicate suppression) on real-world camera imagery
• Own the model-serving pipeline — package models into our NVIDIA Triton ensembles (DALI GPU preprocessing → TensorRT inference  → post-processing), build and deploy TensorRT engines, manage the model repository and no-downtime reloads
• Build and curate datasets — ingestion, labelling, and quality control using FiftyOne(Voxel51) and Label Studio; identify and fix the data problems that actually move model accuracy
• Design evaluation harnesses so model changes are measured, not guessed — regression suites, A/B comparisons, and metrics tied to real detection quality
• Develop LLM/VLM and agentic capabilities — extend our self-hosted VLM/LLM stack(vLLM and similar), build retrieval- and tool-using agents, and integrate them into engineering and product workflows

Qualifications
Must have:
• Strong Python and the modern ML stack — PyTorch, model training and fine-tuning, working in Jupyter / notebook-driven experimentation
• Practical computer vision experience — object detection, working with image data, understanding why models fail in the real world
• Experience taking models to production, not just training them — model serving, optimization, and the gap between offline metrics and live behavior
• Self-starter mindset — you can take an ambiguous accuracy problem, dig into the data, run the experiments, and ship a measurable improvement independently
• Rigorous about evaluation — you care about datasets, ground truth, edge cases, and not fooling yourself with a good-looking number

Nice to have:
• NVIDIA Triton Inference Server, TensorRT, DALI, or comparable GPU model-serving / optimization experience
• Dataset tooling — FiftyOne (Voxel51), Label Studio, or similar curation/annotation platforms
• LLM / VLM experience — self-hosting (vLLM), fine-tuning (LoRA), RAG, or multimodal models
• Agent-building experience — tool-using agents, MCP, or LLM-orchestration frameworks
• MLOps — experiment tracking (CometML/Opik or similar), model registries, reproducible training pipelines
• Exposure to edge/IoT or resource-constrained inference, or to anomaly detection on device telemetry
• Familiarity with NATS / gRPC or other event-driven service communication

Additional Information
Level
Mid-level (2–5+ years of relevant ML engineering experience). We value an engineer who can both improve a model and keep it running in production over a pure researcher or a pure MLOps specialist — depth in the CV/serving stack matters more than breadth across every framework.