AI & Agents Portfolio
Four production AI deployments that bridge the gap between high-level reasoning and low-level execution: industrial computer vision, autonomous greenhouse intelligence, maritime predictive maintenance, and agentic fleet infrastructure. Flip the cards to see the neural levers behind the systems.
Selected Intel
Computer Vision
High-performance visual inspection on the manufacturing floor using Jetson Nano and TensorRT.
Insight: Implemented custom CUDA kernels and a zero-copy memory pipeline to handle 60fps inspection with sub-50ms latency.
Impact: Reduced model memory footprint by 40% while maintaining 99.8% detection accuracy in variable lighting.
Industrial Defect Detection
The Hard Part
Standard object detection models are too heavy for edge silicon. Running a naive YOLOv8 on a Jetson Nano causes thermal throttling and frame drops, missing critical defects on a moving line.
The Approach
We optimized the model using TensorRT INT8 quantization and moved the preprocessing (resizing, normalization) into custom C++ kernels that utilize the Jetson's VIC (Video Image Compositor). This freed up the GPU to focus entirely on inference, reaching a rock-solid 60fps.
Edge Reasoning
LLM-augmented decision engine for multi-zone greenhouse optimization.
Insight: Built a hybrid RAG system that combines real-time sensor data with historical crop growth models for autonomous vent and heating control.
Impact: Optimized CO2 usage by 22% and reduced heating costs by 15% through predictive thermal modeling.
Autonomous Climate Reasoning
The Hard Part
LLMs are great at reasoning but hallucinate physical limits. An AI agent might suggest opening a vent for airflow while ignoring that it's currently raining outside, potentially damaging the crop.
The Approach
We implemented a "Safe-Reasoning" layer. The AI agent generates a high-level strategy, which is then validated against a deterministic rule-engine built in Elixir. If the AI suggests an action that violates physical safety constraints, the system automatically pivots to a "safe-state" fallback.
Predictive Ops
Anomaly detection for maritime engine systems handling 2M+ telemetry signals.
Insight: Designed a distributed anomaly detection pipeline that flags engine wear patterns before they lead to mid-ocean failures.
Impact: Achieved a 3-week lead time on critical component failures, saving an estimated 120K EUR per vessel in avoided downtime.
Predictive Vessel Maintenance
The Hard Part
Ship telemetry is noisy and sparse due to satellite link drops. Standard time-series models fail when 20% of the data points are missing or corrupted by electrical interference.
The Approach
We used a Variational Autoencoder (VAE) to reconstruct missing signals and learn the "normal" manifold of engine behavior. The model runs on the edge gateway, only transmitting "anomaly scores" and supporting snippets of raw data when a threshold is crossed, saving bandwidth.
Agentic Systems
Autonomous coordination for 80+ EV charging stations using agentic negotiation protocols.
Insight: Built a multi-agent system where each charging station acts as a self-interested negotiator for grid capacity.
Impact: Self-balancing grid load that reacts to pricing spikes in under 500ms without a centralized bottlenecks.
Agentic Fleet Infrastructure
The Hard Part
Centralized schedulers are single points of failure. If the main server lags, 80 chargers could simultaneously spike the grid, blowing the depot's main fuse.
The Approach
We treated each charger as an autonomous agent. They use a gossip protocol to "bid" for available power based on their assigned vehicle's priority. This creates an emergent, self-organizing load balancer that is naturally resilient to network partitions.
Neural Trace
Waiting for intelligence reveal
Agentic Infrastructure Notes
When the character triggers a card lever, the neural reasoning reveal is activated, surfacing the design decisions behind sensing intelligence and agentic control.
AI is most effective when it is directly coupled to physical sensors. Intelligence starts at the edge, not the cloud.
Decentralized agents are more resilient than centralized schedulers. Design systems that coordinate via emergent behavior.
AI should suggest strategies, but deterministic rule-engines must enforce physical safety boundaries.
Edge silicon has limits. Quantization, pruning, and custom kernels are not optimizations—they are requirements.
Inference Bench
Designing vector-space retrieval over high-frequency industrial sensor streams for LLM context.
Researching partition-tolerant coordination patterns for decentralized AI agent negotiation.
Optimizing custom transformer models for low-power ARM and Jetson silicon with minimal precision loss.
Personal Philosophy