A3GIS brings together agent infrastructure, evidence-led engineering, computer vision and assistive IoT. The work explores how intelligence can become useful in the real world.
A model can recognise a pattern. The harder question is what that recognition makes possible for someone.
That question runs through A3GIS: sensing the world, making the signal useful, giving actions clear boundaries, and asking for evidence. These are independent projects connected by a shared research direction.
Selected work
Four projects. A shared inquiry.
01 / Flagship agent infrastructure
In development
AEGIS-COGNITION
From intent to evidence.
The core agent infrastructure project: a small Python-facing API, with Rust owning task admission, resource limits, cancellation and evidence transitions.
The architecture makes ownership and replay boundaries explicit. It asks a practical question: when an agent acts, who holds the boundary, and what evidence remains?
Open research question
How could these explicit boundaries support systems that eventually act beyond a software interface?
Active development. Production readiness remains unverified.
A conceptual view of the documented architecture. Rust owns authority and evidence; this is not a live execution.Python API / README excerpt +
from aegis_cognition import Agent
result = Agent(
task="Summarize the supplied research notes"
).run()
A small interface. An explicit authority layer. Formatted from the published example; this page does not execute the runtime.
02 / Assistive IoT
Working prototype
SmartCane
A clearer sense of what’s ahead.
A distance reading matters when it becomes a cue someone can use. SmartCane brings that question into a working ESP32 prototype for people with visual impairments.
The project combines distance sensing, configurable alerts, remote monitoring and an SOS flow. Its purpose gives the engineering a concrete point of reference: a person, a device, and the space around them.
Inside the prototype +
The firmware supports ultrasonic and laser distance sensors, with automatic selection and fallback. Blynk provides app controls, remote monitoring and the SOS workflow.
Configurable reminders and diagnostic reporting make device behaviour easier to inspect. Hardware testing and the next sensing experiments remain separate parts of the work.
Next research question
Could compact, task-focused vision add context to distance sensing while respecting the device’s memory, response time and power limits?
Signal study / Sense → respond
Distance → a useful cue. Illustration of the research question, not a sensor measurement.
Platform
ESP32
Focus
Assistive sensing
Stage
Hardware prototype
Perception study / Model → device02
Illustrated research directionEmbedded deployment is a future goal.
03 / Perception
Collaborative coursework
Computer Vision
From a camera feed to the edge.
The foundation is a collaborative face detection, age and gender estimation project at Vietnamese-German University.
Its documented pipeline combines YOLOv11s-face detection, a MobileNetV3 model and temporal smoothing, with ONNX and OpenVINO paths for desktop inference.
Research direction
Build on this foundation toward task-focused perception on embedded IoT chips, then explore how it could inform robotic systems. The first questions are memory, latency, power and what the model actually needs to recognise.
A3GIS brings together agent infrastructure, engineering methods and applied research.
Public repositories hold the architecture, the implementation approach and the questions still open. The assistive hardware work adds a physical context to that inquiry.
