Notes

Writing on AI in practice.

What I have learned building these systems for real: the failures worth knowing about, the measurements that matter, and the decisions I would make differently.

Retrieval-augmented chatbot answering a question with the source passages it used shown beside the answer

Generative AI

14 Jul 20263 min read

Why your RAG chatbot answers confidently and wrongly

Most retrieval-augmented systems fail at retrieval, not generation. Here is where they break, and the fixes that moved the needle on a bilingual assistant handling real student queries.

RAG
Vector Search
LLM
Evaluation
Chunking

Read

Classification report showing high overall accuracy alongside near-zero recall on the minority class

Machine Learning

28 May 20263 min read

Your model is 97% accurate and completely useless

On imbalanced datasets, accuracy measures how common the majority class is. A walk through what to measure instead, and why SMOTE is not the fix people think it is.

Imbalanced Data
Metrics
SMOTE

Read

Dashboard showing per-conversation token usage and cost for a production LLM assistant

Engineering

9 Apr 20263 min read

Shipping an LLM feature without a surprise invoice

Token costs behave nothing like server costs: they scale with conversation length, not traffic. The caching, routing and truncation decisions that keep a production LLM feature affordable.

LLM
Cost
Caching

Read

Agent execution trace showing the sequence of tool calls taken to answer a data question

Generative AI

2 Mar 20263 min read

When an AI agent is the wrong answer

Agent frameworks make it easy to build something that looks impressive and behaves unpredictably. A practical test for whether your problem needs an agent or a pipeline.

AI Agents
LangGraph
CrewAI

Read

Side-by-side comparison of a base model response and a fine-tuned model response to the same instruction

Generative AI

20 Jan 20263 min read

Fine-tuning or RAG? The question is what you are missing

Fine-tuning teaches behaviour, retrieval supplies facts. Confusing the two is why people fine-tune a model on their documentation and are disappointed by the result.

Fine-tuning
LoRA
QLoRA

Read

Scanned handwritten page beside the transcribed text output with corrected words highlighted

Machine Learning

11 Nov 20252 min read

OCR on real handwriting: what the tutorials leave out

Off-the-shelf OCR reads printed text well and handwriting badly. The preprocessing, segmentation and post-correction steps that took a handwritten transcription system from unusable to useful.

OCR
Computer Vision
OpenCV

Read

Prefer a reader? Subscribe via RSS.

Next step

Have an AI problem worth solving?

Tell me what you are trying to build. I will tell you honestly whether AI is the right tool, what it would take, and how I would approach it.

Start a project Book a call tasrifnurhimel@gmail.com

Typical reply within 24 hours

Working from Dhaka, Bangladesh · remote worldwide