Skip to content

KSKunj ShahAI Engineer & Agent Builder

Work
Services
Pricing
Writing
About

Start a project

Loading…

Kunj Shah

AI engineer in Ahmedabad, India. I build agents, document search and full-stack AI apps, with the tests that prove they work.

Start a project

Home
About
Experience
Education
Projects
Open Source
Skills
Labs
Manifesto
Hackathons
Writing
Contact

GitHub
LinkedIn
X
Product Hunt
Resume
Email

© 2026 Kunj Shah

Cite
Privacy
Terms

Ahmedabadlocal time, updates every minute

Home
Projects
Writing
Experience
Contact

Edge & Vision/October 20, 2026/8 min read

Edge Computer Vision: Bringing AI Vision to Hardware That Doesn't Cost $10,000

Running YOLOv8 on edge devices like NVIDIA Jetson gives you real-time object detection without the cloud costs.

yolov8edge computingcomputer visionjetson

Running YOLOv8 on edge devices like NVIDIA Jetson gives you real-time object detection without the cloud costs.

Related Questions

How do AI agents handle CAPTCHAs?: AI agents use human-in-the-loop checkpoints, session continuity, and site-specific adaptations to gracefully handle CAPTCHAs rather than trying to bypass them.
What's the best edge device for computer vision?: NVIDIA Jetson Nano ($199) offers the best balance of performance and price, capable of 22 FPS with INT8-quantized YOLOv8.
Is OWASP relevant for AI projects?: Yes - AI agents need input validation, output sanitization, session management, and authorization checks at every step, similar to traditional web app security.

1,876 words · first published October 20, 2026

Kunj Shah is an AI engineer in Ahmedabad, India, building agents, retrieval systems, and edge vision that hold up in production. More about the work, or get in touch.

People also ask

How accurate is edge computer vision on real track footage?+

The deployment reached 98.7% detection accuracy on held-out track footage, using YOLOv8 compiled to a custom TensorRT INT8 engine on an NVIDIA Jetson Orin.

Why use INT8 quantization for edge inference?+

INT8 delivers roughly 2x the throughput of FP16, and with quantization-aware training the accuracy you keep is real rather than forcibly truncated into 8-bit precision.

What does edge computer vision still get wrong?+

Extreme low-light, novel defect classes absent from the training data and camera misalignment. Histogram equalization helps with blur and low light but cannot create photons.

Read next

Newer · Oct 2026 · 6 min readYOLOv8 vs TensorRT: From Trained Model to Sub-100ms on Edge Hardware Older · Oct 2026 · 6 min readEdge Computer Vision in the Field: What 98.7% Accuracy Actually Required

More on this

Edge & Vision · Oct 2026 · 6 min readYOLOv8 vs TensorRT: From Trained Model to Sub-100ms on Edge HardwareYOLOv8 vs TensorRT is the wrong fight: model family and inference runtime compose. Inside the PyTorch to INT8 pipeline shipped on Jetson Orin for rail defect detection.
Agents · Nov 2026 · 6 min readHow Much Does an AI Agent Cost in India? (2026 Breakdown)How much does an AI agent cost in India? Fixed-scope projects from ₹25,000, hourly ₹1,500, retainers ₹30,000-75,000/month. The full 2026 breakdown of what drives price.
Agents · Nov 2026 · 9 min readAI Agent Development Guide: From Brief to ProductionHow to plan, build and ship an AI agent that survives production: scope, tool contracts, evals, verification gates and the fixed-scope pilot that de-risks it all.

Working on something like this?

If one of these problems is the one you are trying to solve, say which and I will answer on that directly.