Taking AI beyond experiments,
into real business.
Starting with business problems, connecting models, engineering and client delivery.
From problem to solution
Understand business goals, identify data, compute and field constraints, and design a testable technical approach.
Solution design / PoCFrom model to system
Connect computer vision, retrieval and model capabilities to backends, databases and business workflows.
Computer vision / RAG / BackendFrom validation to delivery
Handle offline environments, edge compute and system integration so the technology works in the field.
Deployment / Edge AIProjects
How I solve problems, through projects.
Industrial parts · Visual recognition and retrieval
Industrial parts recommendation: from visual measurement to a business system
How can specification data support explainable recommendations when product photos are missing?
Equipment verification · VLM extraction
Nameplate extraction: balancing quality, compute and review
Extract structured fields from blurred, reflective or incomplete nameplates while preserving uncertainty.
Driver training · Multi-module vision analysis
Dual-camera driving diagnosis: turning complex judgments into traceable evidence
How can two video streams provide evidence that driving instructors can review?
Railway equipment · Visual inspection
Equipment anomaly detection: from single frames to combined evidence
Combine limited anomaly samples, segmentation and consecutive frames into a testable inspection pipeline.
Edge AI · Model optimization and delivery
Visual localization quantization: diagnosing accuracy loss for edge deployment
After shrinking a model, how do we check that it still retains its localization capability?
Manufacturing and logistics · Video analytics and delivery
Video analytics SaaS: from client PoCs to product decisions
Beyond recognition quality, can clients use and maintain the solution, and can delivery scale?
Industry collaboration · Activity recognition and edge deployment
Machinery activity recognition: reaching the field with limited resources
How can time-series recognition balance accuracy and real-time processing on constrained hardware?
Master’s research · Efficient activity recognition
Sensor selection and model compression: reducing system burden with sparse training
Can training reduce both sensor requirements and model size while retaining recognition quality?
Technical notes
Technical notes are on the way.
I will share project reflections, technical decisions and engineering practice here.