Hasan Goni – Hasan’s Data Science & AI Blog

Categories

All (129)

a2a (1)

advanced (2)

agentic-ai (32)

agents (1)

AI (2)

analysis (4)

anomaly detection (2)

anomaly-detection (4)

application (1)

audio-generation (3)

automation (3)

awk (1)

background-jobs (1)

bash (3)

basics (3)

beginner (1)

Benchmarks & Evaluation (4)

bsub (1)

buffers (1)

career (1)

change (1)

chmod (1)

chown (1)

classical-cv (1)

clipboard (1)

code (4)

Colab (1)

command-line (13)

Computer Vision (1)

computer-vision (79)

configuration (1)

continual-learning (1)

contours (1)

ControlNet (1)

course-notes (3)

cp (1)

cron (1)

customization (1)

data science (2)

data-science (4)

Deep Learning (1)

deep-learning (5)

delete (1)

deployment (1)

depth-estimation (1)

development (2)

diffusion (8)

diffusion-models (2)

dinov2 (1)

distillation (1)

edge-ai (4)

edge-detection (1)

editing (1)

efficiency (1)

environment (1)

evaluation (2)

expert (1)

fastai (2)

feature-engineering (1)

features (1)

few-shot (1)

file-operations (1)

file-transfer (1)

files (1)

find (1)

fine-tuning (2)

Fine-tuning (1)

Flux (1)

foundation-models (2)

GAN (1)

Generative Models (2)

Generative Retrieval (1)

generative-modeling (2)

gnu-parallel (1)

grep (1)

hpc (6)

HuggingFace (2)

image (1)

image-processing (1)

image-registration (5)

inspection (5)

intermediate (1)

job-scheduling (1)

keypoints (1)

langgraph (6)

learning (1)

licensing (1)

linux (1)

LLM (2)

LoRA (7)

machine learning (2)

Machine Learning (1)

machine-learning (1)

macros (1)

matching (1)

mathematics (1)

matplotlib (1)

matrix multiplication (1)

mcp (1)

medical-imaging (1)

mkdir (1)

Model Efficiency (3)

modes (1)

monitoring (1)

morphology (1)

movement (1)

Multimodal Learning (3)

mv (1)

navigation (2)

neural-networks (1)

news (1)

NLP (1)

Nlp (1)

numpy (1)

ood-detection (1)

open-source (10)

opencv (3)

OpenCV (1)

operators (1)

opinion (1)

orchestration (1)

packaging (1)

paper-deep-dive (1)

papers (24)

parallel-computing (2)

parallel-processing (1)

patterns (1)

permissions (1)

philosophy (1)

physics (1)

plugins (1)

precision (1)

productivity (1)

programming (1)

project (1)

python (21)

pytorch (2)

QLoRA (1)

rag (1)

Reasoning (1)

Reasoning & Agents (6)

regex (1)

registers (1)

Reinforcement Learning (1)

remote-access (1)

remote-sensing (3)

replace (1)

Research paper (1)

resources (1)

retrieval (1)

rm (1)

rsync (1)

sam3 (1)

scp (1)

scripting (2)

search (2)

security (3)

sed (1)

segmentation (9)

selection (2)

semiconductor (24)

series (23)

software-factory (1)

sort (1)

splits (1)

ssh (1)

structured-output (1)

supply-chain (1)

survival (1)

Synthetic Data (1)

synthetic-data (13)

tabs (1)

tabular (1)

Text (1)

text (1)

text-editor (1)

text-objects (1)

text-processing (1)

thresholding (1)

tools (2)

Transformers (1)

transformers (1)

tutorial (3)

unix (1)

uv (1)

validation (1)

variables (1)

video-generation (3)

video-segmentation (1)

vim (11)

vision-language (1)

Vision-Language Models (5)

vision-language-models (3)

visual-mode (1)

VLM (14)

weekly-pick (10)

workflow (2)

X-ray (13)

xargs (1)

yank (1)

Computer vision, applied ML, and practical tooling — written up as I use it, mostly in series rather than one-off posts. Start with a series or browse everything below.

Agentic Developer: Zero to Advanced

series

agentic-ai

langgraph

computer-vision

python

A from-scratch path through agentic AI development – tool calling, LangGraph, multi-agent orchestration, RAG, fine-tuning, VLMs, and production hardening – all built around one running use case: X-ray void detection on a synthetic dataset with exact ground truth.

ControlNet-Conditioned Diffusion

computer-vision

synthetic-data

X-ray

diffusion

ControlNet

LoRA

Part 2’s diffusion model learned real defect texture but not real defect shape. This post tests whether that’s a conditioning problem, not a model-scale problem.

Joint Image and Mask Diffusion

computer-vision

synthetic-data

X-ray

diffusion

generative-modeling

Parts 2-4 all treat the defect mask as something supplied from outside. This post inverts that: a small model learns to generate the image and its own mask together, from scratch.

Physics-Based X-ray Simulation

computer-vision

synthetic-data

X-ray

physics

No access to the X-ray equipment that shot our images? Invert the physics on a real image instead of simulating from a CAD model that doesn’t exist.

Flux-Based Mask-Conditioned Generation

computer-vision

synthetic-data

X-ray

diffusion

Flux

LoRA

Parts 2-3 both fine-tuned Stable Diffusion UNets. This post swaps in a genuinely different architecture – a 12B-parameter diffusion transformer – to test whether scale, not conditioning, is what closes the realism gap.

Mask-Conditioned Diffusion Fine-Tuning

computer-vision

synthetic-data

X-ray

diffusion

LoRA

Part 1 got the physics right by calibration. This post asks the mirror-image question: what if a model just learned real defect texture from data instead?

GAN-Based Synthesis

computer-vision

synthetic-data

X-ray

GAN

generative-modeling

A distinct third generative family from the diffusion track: a defect-aware conditional GAN, trained from scratch on 338 real examples – and the honest, textbook way it broke.

SpatialClaw for Inspection

series

agentic-ai

computer-vision

inspection

A step-by-step build of a code-as-action agent for image-only inspection, inspired by NVIDIA’s SpatialClaw: what the paper really does, what to keep and drop, and how to test whether it beats a plain tool-call agent.

X-ray Inspection Video + Sound Generation

computer-vision

video-generation

audio-generation

X-ray

diffusion

A real weld radiograph, a real 19B joint video+audio diffusion model, and a real, honest failure mode: the model keeps the first frame in the X-ray domain and drifts everywhere else.

Video Conditioning vs. Single-Frame Conditioning

computer-vision

video-generation

audio-generation

X-ray

diffusion

Part 1’s negative frame conditioning drifted from X-ray into color photography. A reader question about a real LTX-2 feature – video conditioning – fixes half the problem, and the measured result shows exactly which half.

Local-AI Defect Inspection Series

series

computer-vision

VLM

semiconductor

edge-ai

Testing small, local vision-language models against real industrial defect imagery — semiconductor, casting, wafer — and being honest about where they break.

Agentic AI Software Factory: Inspection System

series

agentic-ai

software-factory

langgraph

Building a factory of AI agents that takes tickets, changes the X-ray inspection pipeline from Agentic Developer: Zero to Advanced, and merges only what passes a gate the agents cannot edit – planned from published evidence, built in small steps, with real runs and honest numbers.

Agentic AI Software Factory: Robust Segmentation Models

series

agentic-ai

segmentation

computer-vision

A factory of AI agents that proposes, runs and judges experiments to build better, more robust segmentation models for X-ray inspection – with a sealed test set, a locked evaluator, and a plain hyperparameter sweep as the honest control. Planned from published evidence, built in small steps.

Agentic Developer: Zero to Advanced, Part 14: Context Engineering Starts With Retrieval

grep, paged reads, offloading, paths instead of payloads, and sub-agents that see one file. Measured on this series’ own archive, and the result is not the one the checklist suggests.

agentic-ai

python

retrieval

evaluation

This post is part of the Agentic Developer: Zero to Advanced series. It stands on its own: it uses the synthetic void data from Part 2 and the reports the pipeline from Part 7…

Paper of the Week

series

papers

computer-vision

VLM

A handful of papers a week, picked and explained for my own semiconductor CV/VLM defect-inspection work — ELI5 first, then the full technical breakdown.

4 Papers Worth Reading — Week of October 23, 2026

papers

weekly-pick

computer-vision

VLM

Vision-Language Models

Model Efficiency

Synthetic Data

This post is part of the Paper of the Week series — a handful of papers a week, picked and explained for my own semiconductor CV/VLM work.

SpatialClaw for Inspection, Part 5: A Real Sandbox, and Trying to Break It

Parts 3 and 4 ran the model’s code as me, on my machine. This part puts it in a locked container, replays 634 recorded cells to prove nothing changed, and throws 15 attacks at it. It holds, with caveats I list.

series

agentic-ai

computer-vision

inspection

security

This post is part of the SpatialClaw for Inspection series. It replaces the unsandboxed kernel of Part 3 and Part 4. All code is my own.

SpatialClaw for Inspection, Part 4: Eyes and Hands for the Loop

Part 3’s agent could only read numbers it printed. I gave it show() and four small tools, ran the same 12 images three ways, then ran everything twice. The honest answer is not the one I hoped for.

series

agentic-ai

computer-vision

inspection

This post is part of the SpatialClaw for Inspection series. It extends the loop from Part 3. All code below is my own, written from the paper’s description.

SpatialClaw for Inspection, Part 3: The Smallest Kernel Loop

About 270 lines of my own code, a local 27B model, 12 real images with exact ground truth, and a result that is honest about how little it proves.

series

agentic-ai

computer-vision

inspection

This post is part of the SpatialClaw for Inspection series. It builds the loop planned in Part 2. Everything in the loop is my own code, written from the description in the…

SpatialClaw for Inspection, Part 2: The Adaptation Plan

What to keep, what to swap, what to drop, what the licences allow, and how I will know whether it was worth it.

series

agentic-ai

computer-vision

inspection

This post is part of the SpatialClaw for Inspection series. It builds on Part 1, the deep dive. Everything here is a plan. Nothing in this post has been built or measured…

SpatialClaw for Inspection, Part 1: What SpatialClaw Actually Does

A deep dive into NVIDIA’s code-as-action agent: the paper, the official code, two pieces I ran, and where it is weaker than the headline.

series

agentic-ai

computer-vision

paper-deep-dive

This post is part of the SpatialClaw for Inspection series. The next post, Part 2, turns what we learn here into a plan for image-only inspection.

A Synthetic Void Dataset With Exact Masks, Built From Fashion-MNIST

Agentic AI Software Factory: Robust Segmentation Models, Part 1. A dataset I am free to share, where every defect mask is exact because the defect is drawn, not labelled.

segmentation

synthetic-data

computer-vision

agentic-ai

Update, 2026-10-19: there is a v2. Looking at contact sheets of all 4,000 images, I found that v1 has unlabelled look-alike patches (64% of images) and labelled voids that…

Agentic Developer: Zero to Advanced, Part 8: Tool Design

Plain code, a small typed model, or a language model? Three kinds of tool, measured head to head on two real calls from the Part 7 pipeline.

agentic-ai

python

computer-vision

evaluation

This post is part of the Agentic Developer: Zero to Advanced series. It follows on from Part 7 – the inspection pipeline, and it is the fourth post of Arc 2, the agent loop.

Agentic Developer: Zero to Advanced, Part 6: Building the Graph

LangGraph basics: state, nodes, edges. The Part 5 loop, redrawn so you can see it, pause it and test it.

agentic-ai

langgraph

python

computer-vision

This post is part of the Agentic Developer: Zero to Advanced series. It follows on from Part 5 – what makes it an agent, and it is the second post of Arc 2, the agent loop.

Agentic Developer: Zero to Advanced, Part 7: The Inspection Pipeline

Classify, check the rules, decide, report. A real LangGraph graph where the code owns the order and the model owns two judgments.

agentic-ai

langgraph

python

computer-vision

This post is part of the Agentic Developer: Zero to Advanced series. It follows on from Part 6 – building the graph, and it is the third post of Arc 2, the agent loop.

5 Papers Worth Reading — Week of October 9, 2026

papers

weekly-pick

computer-vision

VLM

Reasoning & Agents

Generative Models

This post is part of the Paper of the Week series — a handful of papers a week, picked and explained for my own semiconductor CV/VLM work.

Agentic Developer: Zero to Advanced, Part 5: What Makes It an Agent

Planning, tool use, memory, control flow. Written by hand in one loop, then scored against exact ground truth.

agentic-ai

python

computer-vision

This post is part of the Agentic Developer: Zero to Advanced series. It follows on from Part 4 – prompting and structured output. It opens Arc 2, the agent loop.

5 Papers Worth Reading — Week of October 2, 2026

papers

weekly-pick

computer-vision

VLM

Vision-Language Models

Reasoning & Agents

Benchmarks & Evaluation

Multimodal Learning

This post is part of the Paper of the Week series — a handful of papers a week, picked and explained for my own semiconductor CV/VLM work.

FleXray’s Synthetic X-Ray Engine — What Actually Transfers to Universal Void Detection

MIT/MGH’s whole-body X-ray segmentation model doesn’t touch industrial X-ray at all – but its data engine has real, specific lessons for scarce-defect synthetic data, and one big one that doesn’t transfer

papers

computer-vision

segmentation

synthetic-data

semiconductor

This post is part of the Paper of the Week series — an applicability audit of the FleXray data engine for void detection, not a full model reimplementation.

Can Qwen-Image-2.1 Draw You a Rare Defect? Mask-Conditioned Editing, Tested Honestly on Real X-Rays

Qwen/Qwen-Image-2.1 (7B DiT) tried as a new tool for this blog’s own rare-defect synthetic-data problem – real mask edits, real diversity/severity sweeps, real failure modes, on real GDXray Castings X-rays

papers

computer-vision

diffusion-models

synthetic-data

semiconductor

Images removed. This post’s figures showed frames from GDXray or outputs of models trained on them. GDXray is free for research and education only, and redistribution is not…

Agentic Developer: Zero to Advanced, Part 4: Prompting and Structured Output

Getting reliable JSON/structured responses from the LLM about defect type, severity, confidence

agentic-ai

structured-output

computer-vision

python

This post is part of the Agentic Developer: Zero to Advanced series. It follows on from Part 3 – your first tool-calling LLM.

Agentic Developer: Zero to Advanced, Part 3: Your First Tool-Calling LLM

Wrap the Part 2 classical detector as a callable tool: one LLM request, real structured output, plain-English explanation

agentic-ai

computer-vision

synthetic-data

python

This post is part of the Agentic Developer: Zero to Advanced series. It follows on from Part 2 — the dataset and the classical baseline.

Reimplementing Marigold V2 — One Forward Pass, and the Fine-Tuned Checkpoint That Made Things Worse

arXiv:2609.08084 — the real huawei-bayerlab/marigold-v2-0 LoRA vs. the raw, un-fine-tuned Qwen-Image-Edit-2509 DiT it sits on, both run on real GDXray+ Castings X-rays from this blog’s own open-source defect dataset

papers

computer-vision

diffusion-models

depth-estimation

semiconductor

Images removed. This post’s figures showed frames from GDXray or outputs of models trained on them. GDXray is free for research and education only, and redistribution is not…

5 Papers Worth Reading — Week of September 11, 2026

papers

weekly-pick

computer-vision

VLM

Reasoning & Agents

Vision-Language Models

Benchmarks & Evaluation

Model Efficiency

Generative Models

This post is part of the Paper of the Week series — a handful of papers a week, picked and explained for my own semiconductor CV/VLM work.

Reimplementing SAM3-O2D2 — When a Detector Lies, Make It Prove the Claim, With Real Code

arXiv:2609.08281 — real facebook/sam3 + a real torchvision detector as mutual witnesses, tested on a real natural image, a constructed OOD probe, and real semiconductor X-ray images

papers

computer-vision

ood-detection

sam3

semiconductor

This post is part of the Paper of the Week series – same standard as every deep-dive here: real code, real data, a real measured result, no pseudocode.

Agentic Architecture: Supply Chain

series

agentic-ai

langgraph

supply-chain

The same planning, tool-use, and reasoning architecture from Agentic Developer: Zero to Advanced, re-applied to a demand/delivery anomaly investigation agent – proof the pattern is domain-agnostic, not X-ray-specific.

Agentic Developer: Zero to Advanced, Part 2: The Dataset and the Baseline Model

A plain classical-CV detector, no agent anywhere in sight, scored honestly against exact ground truth on a synthetic void dataset

agentic-ai

computer-vision

synthetic-data

python

This post is part of the Agentic Developer: Zero to Advanced series. It follows on from Part 1 – why agentic AI, why X-ray defect detection.

Agentic Developer: Zero to Advanced, Part 1: Why Agentic AI, Why X-ray Defect Detection

Setting up the running example the rest of this series builds on

agentic-ai

langgraph

computer-vision

synthetic-data

python

This post is part of the Agentic Developer: Zero to Advanced series.

Synthetic Data Series

series

computer-vision

synthetic-data

X-ray

semiconductor

Generating synthetic X-ray defect images for industrial inspection when real defective examples are scarce – physics-based simulation, generative fine-tuning, GANs, and the classic cut-paste baseline, measured against each other.

Reimplementing ENEAS — Catching a Tracker Before It Grabs the Wrong Fish, With Real Code

arXiv:2609.03756 — real SAM2 tracking-by-detection, a real semantic verification layer (DINOv2 + Qwen2-VL), on a real DAVIS video with 6 look-alike distractors

papers

computer-vision

video-segmentation

vision-language-models

open-source

This post is part of the Paper of the Week series – same standard as every deep-dive here: real code, real data, a real measured result, no pseudocode.

Reimplementing NC-TFAD — Fixed Prototypes vs. a Naive Baseline, Streamed Honestly

arXiv:2609.03406 — real simplex-ETF prototypes for task-free continual anomaly detection, measured against a naive drifting-mean baseline on all 15 real MVTec-AD categories, no announced task boundaries

papers

computer-vision

anomaly-detection

continual-learning

open-source

This post is part of the Paper of the Week series – real code, real data, real measured numbers, reported honestly even when a paper’s claimed advantage doesn’t show up.

Reimplementing PL-SCEA — Take Two, After a GitHub Issue Told Me I Got It Wrong

arXiv:2609.03655 — a corrected reimplementation (RADIOv3-B, γ = 2, the real attention hook), run on two data protocols at image and pixel level, and an honest account of what does and doesn’t reproduce

papers

computer-vision

anomaly-detection

open-source

This post is part of the Paper of the Week series – real code, real data, real measured numbers, reported honestly even when they don’t flatter the paper.

5 Papers Worth Reading — Week of September 4, 2026

papers

weekly-pick

computer-vision

VLM

Vision-Language Models

Generative Retrieval

Reasoning & Agents

Multimodal Learning

This post is part of the Paper of the Week series — a handful of papers a week, picked and explained for my own semiconductor CV/VLM work.

Head-to-Head, Licensing, and the Recommendation

computer-vision

image-registration

semiconductor

licensing

Every registration method’s measured accuracy side by side, a license table checking code and pretrained weights separately, and a concrete recommendation for semiconductor inspection work.

Deep Learning: Deformable Registration

computer-vision

image-registration

deep-learning

semiconductor

A from-scratch VoxelMorph-style network – and an honest first failure that turns out to be a real, well-known lesson: deformable networks need affine pre-alignment, not a raw drifted image.

3 Papers Worth Reading — Week of September 3, 2026

papers

weekly-pick

computer-vision

remote-sensing

anomaly-detection

segmentation

This post is part of the Paper of the Week series — a handful of papers a week, picked and explained for my own semiconductor CV/VLM work.

Reimplementing DPA — Zero-Shot Anomaly Generation by Transfer, With Real Code

arXiv:2609.02075 — a real SD1.5-inpainting LoRA fine-tune on real MVTec-AD data, with a genuine held-out zero-shot test against real ground truth

papers

computer-vision

anomaly-detection

diffusion

open-source

This post is part of the Paper of the Week series. It follows up on the digest write-up of this same paper with a full, real reimplementation.

Reimplementing Lightweight VLM Adaptation for Multispectral + SAR, With Real Code

arXiv:2609.02187 — a real Qwen2-VL-2B fine-tuned via its own multi-image interface, LoRA on real So2Sat Sentinel-1/Sentinel-2 data, ELI5 to full technical breakdown

papers

computer-vision

remote-sensing

vision-language-models

open-source

This post is part of the Paper of the Week series. It follows up on the digest write-up of this same paper with a full, real reimplementation – same standard as every…

Reimplementing InstEditSeg — Segmentation as Instructed Image Editing, With Real Code

arXiv:2609.02004 — a real SD1.5 U-Net + DINO feature guidance, LoRA fine-tuned on real Kvasir-SEG polyp images, ELI5 to full technical breakdown

papers

computer-vision

segmentation

diffusion

medical-imaging

open-source

InstEditSeg: Instruction-Driven Image Editing for Polyp and Skin Lesion Segmentation — Ziquan Liu, Zhewei Zhu, Xuyang Shi (arXiv:2609.02004)

Reimplementing Point-Supervised Change Detection With Real SAM2, With Real Code

arXiv:2609.02171 — real SAM2 point-prompted pseudo-labels on real LEVIR-CD data, a real refinement CNN, and real teacher-student self-training

papers

computer-vision

segmentation

remote-sensing

open-source

This post is part of the Paper of the Week series. It follows up on the digest write-up of this same paper with a full, real reimplementation.

Video & Sound Generation Series

series

computer-vision

video-generation

audio-generation

X-ray

Applying joint video+audio generation models to real industrial-inspection X-ray imagery – what a general-purpose model can and can’t do when handed one real radiograph and asked to imagine motion and sound around it.

Deep Learning: Affine Regression and Deep Feature Matching

computer-vision

image-registration

deep-learning

semiconductor

A from-scratch CNN trained unsupervised to predict an affine transform, and LoFTR as the commercially-licensed deep alternative to SIFT – with a surprising result about which one actually wins.

The Benchmark and Classical Methods

computer-vision

image-registration

semiconductor

classical-cv

A synthetic die image with known ground-truth drift, then phase correlation, ECC, and SIFT+RANSAC put to the test against it – including a real, explained failure mode.

Image Registration Series

series

computer-vision

image-registration

deep-learning

semiconductor

Aligning drifted semiconductor inspection images back to a golden reference – classical methods (phase correlation, ECC, SIFT+RANSAC) vs. deep learning (a from-scratch affine CNN, LoFTR deep feature matching, a VoxelMorph-style deformable network), scored against a synthetic ground-truth benchmark, plus which methods are actually licensed for company use.

5 Papers Worth Reading — Week of August 28, 2026

papers

weekly-pick

computer-vision

VLM

Reasoning & Agents

Benchmarks & Evaluation

Multimodal Learning

This post is part of the Paper of the Week series — a handful of papers a week, picked and explained for my own semiconductor CV/VLM work.

Reimplementing RAU — VLMs Pointing SAM2 at the Right Region, With Real Code

arXiv:2509.22404 — a real Qwen2.5-VL + real SAM2 fusion, trained on real CAMUS cardiac ultrasound, ELI5 to full technical breakdown

papers

computer-vision

segmentation

vision-language-models

open-source

RAU: Reference-based Anatomical Understanding with Vision Language Models — Yiwei Li, Yikang Liu, Jiaqi Guo, Lin Zhao, Zheyuan Zhang, Xiao Chen, Boris Mailhe, Ankush…

Online Course Notes

series

course-notes

Consolidated notes from online courses I work through — lesson summaries, notebook details, and glossaries, kept as a reference rather than scattered across scratch files.

Reimplementing HPMA — Frozen Prototype Memory for SAM3, With Real Code

arXiv:2608.24541 — wired into the real facebook/sam3 model, trained on real EndoVis2017 data, no pseudocode

papers

computer-vision

segmentation

open-source

Hierarchical Prototype-Memory Adaptation of SAM for Surgical Instrument Segmentation — Xinning Yao, Jingjing Wang, Jinghua Yue, Xiaoyan Luo, Fugen Zhou, Bo Liu…

Building Adaptive AI Agents — Course Notes

Consolidated lesson summaries, notebook details, and glossary

agentic-ai

course-notes

This post is part of the Online Course Notes series — consolidated notes from online courses I work through. Want to actually build this instead of just reading about it?…

Building Adaptive AI Agents — Hands-On, Open-Source Edition

Real code, real output, no paid services — reproducing all three course lessons on this machine

agentic-ai

course-notes

open-source

This post is part of the Online Course Notes series, and is the hands-on companion to Building Adaptive AI Agents — Course Notes. That post is the summary; this one is the…

1 Paper Worth Reading — FlowEvo, Week of August 25, 2026

papers

weekly-pick

agents

Reasoning & Agents

This post is part of the Paper of the Week series — a handful of papers a week, picked and explained for my own semiconductor CV/VLM work.

Computer Vision Foundations Series

series

computer-vision

opencv

pytorch

foundation-models

From Pixels to Foundation Models – A Gentle Journey for Beginners. Learn computer vision the fastai way: start with working code, build intuition, then dive deeper.

Command Line Mastery for HPC Series

series

command-line

hpc

linux

bash

parallel-computing

From Windows Explorer to Unix HPC Warrior – Master the essential command line skills for high-performance computing and massive file operations. Learn the Pareto 20% that gives you 80% of the power.

Permission Wizardry

command-line

permissions

chmod

chown

security

Master the ancient art of Unix permissions. Learn to control who can read, write, and execute your files with the precision of a digital locksmith. Essential knowledge for secure HPC environments.

SSH and Remote Access

command-line

ssh

remote-access

security

hpc

Master the art of secure remote connections. Learn to work on distant HPC clusters as if they were sitting on your desk, with the security of a digital fortress and the convenience of local access.

Text Processing Power

command-line

text-processing

grep

awk

sed

sort

Unlock the secrets hidden in your data files. Learn to search, filter, sort, and transform text like a digital archaeologist, extracting insights from millions of lines with surgical precision.

Your Digital Compass

command-line

navigation

basics

unix

Master basic navigation in Unix systems. Learn to find your way around directories like a seasoned explorer, understand where you are, and move confidently through the file system.

File Manipulation Magic

command-line

file-operations

cp

mv

mkdir

rm

Master the art of creating, copying, moving, and organizing files like a digital wizard. Learn to handle thousands of files efficiently with commands that make Windows Explorer look like a toy.

Bash Scripting Essentials

command-line

bash

scripting

automation

hpc

Transform from manual command execution to automation wizard. Learn to create robust bash scripts that handle repetitive HPC tasks, process multiple files, and create reliable workflows that work even at 3 AM.

File Detective Skills

command-line

find

search

hpc

Master the art of finding files among millions. Learn to use find, locate, and grep to track down specific files faster than any GUI could ever manage. Essential skills for HPC environments.

HPC Job Scheduling with bsub

command-line

hpc

bsub

job-scheduling

parallel-computing

Master the art of submitting and managing computational jobs on HPC clusters using bsub. Learn to orchestrate hundreds of parallel jobs, monitor resources, and handle complex workflows like a professional researcher.

File Transfer Mastery

command-line

file-transfer

rsync

scp

hpc

Master the art of moving millions of files efficiently between systems. Learn rsync, scp, and advanced techniques for transferring massive datasets from Windows shares to HPC Unix servers without losing your sanity.

Environment Mastery

command-line

environment

bash

variables

configuration

Master your digital workspace. Learn to customize your shell environment, manage variables, and create a personalized command line experience that adapts to your workflow like a perfectly tailored laboratory.

Automation Excellence

command-line

automation

cron

monitoring

orchestration

Master the pinnacle of command line expertise: complete automation. Learn to create sophisticated systems that think, adapt, and execute complex workflows without human intervention.

Parallel Processing Power

command-line

parallel-processing

xargs

gnu-parallel

background-jobs

Multiply your computational power by mastering parallel processing. Learn to orchestrate multiple processes simultaneously, turning your single-threaded workflows into high-performance parallel operations.

The Magic of Modes: Understanding Vim’s Secret Language

vim

modes

philosophy

intermediate

Discover why Vim’s modal editing is actually genius. Learn to think in Vim’s language and unlock the power of modal text editing.

Vim Survival Guide: Don’t Panic!

vim

beginner

survival

basics

Your first steps into Vim without losing your mind. Learn the absolute basics to survive your first Vim encounter and actually get work done.

Text Objects: The Secret Weapon of Vim Masters

vim

text-objects

selection

advanced

Discover Vim’s most powerful feature - text objects. Learn to select exactly what you want with surgical precision and transform your editing workflow.

The Other 80%: Advanced Vim Mastery

vim

advanced

plugins

customization

expert

Dive into the advanced features that make Vim truly powerful. Explore plugins, customization, and expert techniques for the dedicated Vim enthusiast.

Search and Replace Mastery: Find and Transform Anything Instantly

vim

search

replace

regex

patterns

Master Vim’s powerful search and replace capabilities. Learn to find any pattern and transform text across your entire file with surgical precision.

Registers and Macros: Copy-Paste on Steroids and Automation Magic

vim

registers

macros

automation

clipboard

Master Vim’s advanced clipboard system and powerful macro recording. Learn to manage multiple clipboards and automate repetitive tasks with incredible efficiency.

Navigate Like a Pro: Movement Mastery in Vim

vim

movement

navigation

efficiency

Master the art of moving through text at the speed of thought. Learn the 20% of movement commands that provide 80% of navigation power.

File Management and Buffers: Mastering Multi-File Workflows in Vim

vim

buffers

files

tabs

splits

workflow

Learn to work with multiple files, tabs, splits, and buffers in Vim. Become unstoppable by mastering the art of multi-file editing and navigation.

Visual Mode: See What You’re Doing

vim

visual-mode

selection

precision

Master Vim’s visual selection modes. Learn to select text with precision, see exactly what you’re editing, and apply operations with confidence.

Delete, Change, Copy Like a Wizard: The Holy Trinity of Vim Editing

vim

editing

delete

change

yank

operators

Master the three most powerful editing operations in Vim. Learn to delete, change, and yank text with precision and speed that will make you feel like a coding wizard.

Vim Mastery Series

series

vim

text-editor

productivity

programming

From Text Editor Fear to Vim Ninja – Master the 20% of Vim that gives you 80% of the power. Transform your coding workflow with the world’s most powerful text editor.

5 Papers Worth Reading — Week of August 21, 2026

papers

weekly-pick

computer-vision

VLM

Model Efficiency

Vision-Language Models

Benchmarks & Evaluation

Reinforcement Learning

This post is part of the Paper of the Week series — a handful of papers a week, picked and explained for my own semiconductor CV/VLM work.

Agentic Development Series

series

agentic-ai

computer-vision

semiconductor

Rebuilding DeepLearning.AI’s Agentic AI course (Andrew Ng) around real semiconductor X-ray/AOI imaging work, then going four posts further into production hardening, A2A interoperability, distillation, and retrieval — none of which the course covers.

Agentic AI for Semiconductor Imaging, Part 7: Agent-to-Agent Interoperability with A2A

When agents built by different teams and frameworks need to talk to each other

agentic-ai

computer-vision

semiconductor

python

a2a

This post is part of the Agentic Development Series. It follows on from Part 6 — advanced agentic AI development.

Agentic AI for Semiconductor Imaging, Part 9: Retrieval, Grounding Your Agent in Real Knowledge

Two different retrieval problems: similar past images, and relevant text knowledge, plus a defect-level refinement

agentic-ai

computer-vision

semiconductor

python

rag

This post is part of the Agentic Development Series. It follows on from Part 8 — fine-tuning and distilling your own agent, and closes out the series.

Agentic AI for Semiconductor Imaging, Part 2: Reflection

Teaching an agent to critique its own segmentation output

agentic-ai

computer-vision

semiconductor

python

This post is part of the Agentic Development Series. It follows on from Part 1 — what agentic AI actually is.

Agentic AI for Semiconductor Imaging, Part 3: Tool Use and MCP

From function calling to a real MCP server Copilot and Cursor can use

agentic-ai

computer-vision

semiconductor

python

mcp

This post is part of the Agentic Development Series. It follows on from Part 2 — the reflection agent.

Agentic AI for Semiconductor Imaging, Part 1: What Agentic AI Actually Is

Applications, the four design patterns, and setting up aisuite

agentic-ai

computer-vision

semiconductor

python

This post is part of the Agentic Development Series.

Agentic AI for Semiconductor Imaging, Part 5: Planning and Multi-Agent Systems

From a single agent to a team, in two very different communication styles

agentic-ai

computer-vision

semiconductor

python

This post is part of the Agentic Development Series. It follows on from Part 4 — evaluations and error analysis.

Agentic AI for Semiconductor Imaging, Part 8: Fine-Tuning and Distilling Your Own Agent

Turning a large model’s agent behavior into a small model you can run cheaply, forever

agentic-ai

computer-vision

semiconductor

python

distillation

This post is part of the Agentic Development Series. It follows on from Part 7 — agent-to-agent interoperability with A2A.

Agentic AI for Semiconductor Imaging, Part 6: Advanced Agentic AI Development

Production hardening, memory, security, testing, deployment, observability, and safe iteration

agentic-ai

computer-vision

semiconductor

python

This post is part of the Agentic Development Series. It follows on from Part 5 — planning and multi-agent systems, and closes out the course’s own five modules before the…

Agentic AI for Semiconductor Imaging, Part 4: Evaluations and Error Analysis

The component-tagging method for finding out what is actually broken

agentic-ai

computer-vision

semiconductor

python

This post is part of the Agentic Development Series. It follows on from Part 3 — tool use and MCP.

4 Papers Worth Reading — Week of August 15, 2026

papers

weekly-pick

computer-vision

VLM

This post is part of the Paper of the Week series — a handful of papers a week, picked and explained for my own semiconductor CV/VLM work.

Fine-Tuning LFM2.5-VL-3B for X-ray Defect Detection — Does the Grounding Upgrade Actually Help?

computer-vision

VLM

fine-tuning

LoRA

semiconductor

X-ray

edge-ai

Liquid AI’s new LFM2.5-VL-3B claims improved grounding over LFM2-VL-450M. I LoRA fine-tuned it on the same GDXray X-ray defects and eval set as post 3 to find out — detection improved again, and localization still didn’t work, but it broke in a completely different, more interesting way.

Can a Small Local VLM Catch Its Own Auto-Labeling Mistakes? Testing LFM2.5-VL-3B as a Judge

computer-vision

VLM

LoRA

semiconductor

X-ray

edge-ai

few-shot

Instead of generating labels, I asked LFM2.5-VL-3B to grade them — zero-shot vs. few-shot, on synthetically corrupted GDXray labels. Few-shot nearly doubled recall overall, and completely failed on the one error type that matters most. I show exactly why.

Fine-Tuning LFM2-VL-450M for X-ray Defect Detection and Localization

computer-vision

VLM

fine-tuning

LoRA

semiconductor

X-ray

edge-ai

LoRA fine-tuning a 450M local VLM on real GDXray casting X-ray images — detection got measurably better, localization never worked, and I show the raw predictions that prove it.

Data Science Steps Series

series

data-science

A practical guide to data science following FastAI’s top-down learning approach

VLM Series

series

vision-language

VLM

A practical guide to Vision-Language Models (VLMs) with a focus on open-source models and hands-on PyTorch/HuggingFace workflows

Anomaly Detection Series

data science

machine learning

anomaly detection

A series exploring anomaly detection techniques with fun, accessible explanations.

Finding the Oddballs: A Not-So-Technical Guide to Anomaly Detection(Density Estimation and Robustness)

data science

machine learning

tutorial

anomaly detection

A fun, accessible explanation of anomaly detection using density estimation techniques

Fine-Tuning Qwen3-14B with Unsloth: A Practical Guide for Data Scientists

LLM

Fine-tuning

HuggingFace

LoRA

QLoRA

This post is part of the VLM Series. Feedback and questions are welcome!

Hands-On with Qwen3-14B: A Reasoning-Centric LLM for Data Scientists

AI

LLM

Reasoning

NLP

HuggingFace

Colab

Transformers

As data scientists, we constantly move between raw data and real-world decisions. Whether you’re explaining anomalies, generating insights, or deploying models, reasoning is…

Image Segmentation: Dividing and Conquering

computer-vision

segmentation

morphology

thresholding

Imagine you’re looking at a crowded photo and trying to identify individual people. Your brain automatically separates the image into distinct regions—faces, bodies…

Modern Vision Models: CNNs, Vision Transformers, and DINOv2

computer-vision

transformers

foundation-models

dinov2

Remember when we thought AlexNet was revolutionary in 2012? That was just the beginning! In the past decade, computer vision has evolved at breakneck speed:

Your First CV Project: Putting It All Together

computer-vision

project

application

deployment

You’ve learned about pixels, mastered OpenCV, detected patterns, matched features, and explored cutting-edge deep learning models. Now comes the exciting part: building a…

Why Deep Learning? When Classical Methods Hit the Wall

computer-vision

deep-learning

neural-networks

pytorch

Picture this: It’s 2010, and computer vision researchers are frustrated. They’ve spent decades perfecting edge detection, feature matching, and object recognition. Their…

OpenCV Essentials: Your First Computer Vision Toolkit

computer-vision

opencv

image-processing

Imagine you’re a carpenter, and someone hands you the ultimate toolbox—every tool you could ever need, perfectly organized and ready to use. That’s OpenCV for computer vision!

Finding Patterns: Edges, Contours, and Shapes

computer-vision

edge-detection

contours

opencv

Imagine you’re a detective looking at a crime scene photo. Your brain instantly picks out: - The outline of a footprint in the mud - The edge of a broken window - The shape…

Where to Go Next: Your Computer Vision Journey

computer-vision

career

learning

resources

🎉 Congratulations! You’ve just completed an incredible journey through the world of computer vision. Let’s take a moment to appreciate how far you’ve come:

Feature Magic: What Makes Images Unique

computer-vision

features

keypoints

matching

Imagine you’re doing a 1000-piece jigsaw puzzle. How do you know which pieces fit together? You look for unique features—distinctive colors, patterns, corners, and edges…

Images as Data: How Computers See the World

computer-vision

numpy

matplotlib

basics

Remember when you first learned that music could be stored as numbers? Or that your favorite movie is just a sequence of 1s and 0s? Well, prepare for another “mind-blown”…

Why Computer Vision? Teaching a Robot to See

Computer Vision

AI

Machine Learning

OpenCV

Deep Learning

Imagine you’re teaching a friend to recognize your cat in photos. You’d point out the whiskers, the pointed ears, maybe that distinctive patch of white fur. Now imagine…

UV: Python Packaging, Reimagined (and Lightning Fast)

python

packaging

development

tools

uv

scripting

workflow

opinion

If you’ve spent any time wrangling Python environments, you know the drill: pip for installs, venv for isolation, pip-tools for reproducibility, pipx for CLI tools, and maybe …

An Expert Storyteller

Matrix Multiplication

fastai

matrix multiplication

Matrix Multiplication from fastai course 2022

Tools Series

series

tools

python

development

A series exploring modern Python development tools for productivity, reproducibility, and speed.

Understanding Matrix Multiplication from FastAI

A Deep Dive into Neural Network Fundamentals

deep-learning

mathematics

fastai

This post is part of our deep learning foundations series. You might also be interested in: - Data Science Steps Series - Feature Preprocessing

How to Use Nougat to Read Scientific Paper

Research paper working

Nlp

Research paper

Text

Nougat is a Visual Transformer model designed to enhance accessibility of scientific knowledge in the digital age. It converts scientific documents from PDFs into a markup…

Welcome To My Data Science & Machine Learning Blog

news

data-science

Welcome to my technical blog where I share insights about Data Science, Machine Learning, and Deep Learning!

Data Science Steps to Follow - 06

Validation Strategy and Model Evaluation

data-science

machine-learning

validation

Learn about different validation strategies and how to properly evaluate your models

Data Science Steps to Follow - 05

Feature Extraction from Image

code

image

analysis

This is Part 5 of a 6-part series on data science fundamentals:

Data Science Steps to Follow - 03

Exploring Anonymized data

code

tabular

analysis

This is Part 3 of a 6-part series on data science fundamentals:

Data Science Steps to Follow - 04

Feature Extraction from Text

code

text

analysis

This is Part 3 of a 6-part series on data science fundamentals:

Data Science Steps to Follow - 02

Feature Preprocessing and Generation

data-science

feature-engineering

tutorial

Learn essential techniques for preparing and engineering features in your data science projects.

Data Science Steps to Follow - 01

Exploratory Data Analysis Fundamentals

code

analysis

tutorial

This is Part 1 of a 6-part series on data science fundamentals:

---
title: "Hasan Goni"
subtitle: "Notes on computer vision, machine learning, and the tools behind them"
listing:
  contents: posts
  sort: "date desc"
  type: default
  categories: true
  sort-ui: false
  filter-ui: false
page-layout: full
title-block-banner: true
---

Computer vision, applied ML, and practical tooling — written up as I use it, mostly in series
rather than one-off posts. Start with a [series](posts/series/computer-vision-foundations.qmd)
or browse everything below.