Field notes
Building at the terminal
A running log of applied AI and backend work — data pipelines, inference services, and the boring infrastructure that keeps them alive.
40+ entries · updated weekly
Ann Naser Nabil — AI/ML software engineer
From the lab bench to production.
The open lab notebook of Ann Naser Nabil — an AI/ML software engineer: applied research, systems that stay up, and the craft of shipping software in the open.
~/ann-naser-nabil — zsh
$ whoami
Ann Naser Nabil — AI/ML software engineer
$ work --focus
information retrieval · AI automation
$ project --status
doshomikielts.com — building
$ git log --oneline -3
BENI Global 10 — 522K articles · 10 languages
CB-SentiLex — central bank stance benchmark
RAG playbook — 34 projects · 10 concept tracks
$ deploy --status
✓ live at nabil.dev.bd
$
boot complete — static export · 0 runtime deps
Work
A lab bench, not a showcase
A running record of systems, experiments and the occasional yak shave — written where the work happens.
Systems
Backend systems engineered to stay boring
Databases, queues and services designed so that midnight pages never happen. The goal is boring: it means everything is working.
ML
Applied ML, end to end
From data cleaning to production inference, with the trade-offs written down.
15
public repositories
4
NLP benchmarks & datasets
About
An open lab notebook
What this site is, what it publishes, and why the goal is boring.
The open lab notebook of Ann Naser Nabil, an AI/ML software engineer. Everything here is written where the work happens — at the same terminal, in the same repos, on the same infrastructure the lab runs on.
What gets published is the transcript of work that actually ran: field notes from applied ML, post-mortems from production, and tutorials that show the whole path — trade-offs, dead ends and all.
The philosophy is straightforward: boring is working. Databases, queues and services are engineered so that midnight pages never happen — the infrastructure is unremarkable on purpose.
$ currently building: applied ML in production, and the infrastructure to keep it boring
Skills
The skill matrix
Three benches — AI/ML, backend, systems — and the tools that live on each.
PythonPyTorchTransformersRAG pipelinesLoRA fine-tuning
Writing
Notes from the lab bench
Field reports, post-mortems and tutorials — every post is a transcript of work that actually ran.
Research
BENI Global 10: economic narratives across 10 languages
A multilingual economic narrative corpus — 522,397 economically-relevant articles in 10 languages across 7 language families, filtered from 2.8M raw documents.
Bangla NLPMultilingualpublished
Research
CB-SentiLex: an auditable weak-supervision framework for central bank stance
Stance detection over Bangladesh Bank monetary policy statements (2006–2026): 7,432 classified sentences, a curated lexicon and a trained DistilBERT checkpoint, released with a DOI.
Bangla NLPWeak supervisionDOI
Tutorial
RAG the way it is actually engineered
34 project cards and 10 concept tracks: every lab swaps one part of the RAG pipeline — loading, chunking, retrieval, reranking, evaluation — and measures what changed.
RAGLLMsnotebooks
Projects
Selected work
A few systems and experiments shipped from the lab bench.
FinFact-BD
A benchmark for Bengali financial misinformation detection, accepted at the COLING 2026 FinNLP workshop.
PythonTransformersBangla NLP
view projectCB-SentiLex
Bangladesh Bank monetary policy stance detection benchmark, built with a weak-supervision framework.
PythonWeak supervisionBangla NLP
view projectRAG Playbook
A hands-on RAG portfolio: 34 project cards and 10 concept tracks — every lab swaps one pipeline component and measures what changed.
JupyterRAGLLMs
view projectTracehouse
A local-first terminal intelligence platform.
PythonTerminalLocal-first
view projectContact
Let's build something boring that stays up.
For collaboration, questions, or a good argument about infrastructure.