name: Junaid Aslam
role: AI Engineer · Low-Resource NLP · Open Source
built: Pakistan's first Pashto LLM
languages: [ پښتو, English, اردو, Python ]
eos_token: never emitted
state: decodingEvery tutorial builds a sentiment classifier on IMDB. Every portfolio has the same Titanic notebook. I looked at that and felt something close to physical pain.
Meanwhile: Pashto. Forty million speakers. Poetry older than most nations. On HuggingFace — silence. Tokenizers gagging on ښ ږ ډ ړ ټ ڼ ې ۍ. Datasets that were three scraped news sites in a trench coat. No benchmark. No baseline. Nobody coming.
@@ speakers, in millions — and who the models were built for @@
English ████████████████████████████████████████ ~1500
Hindi ████████████████ ~600
Arabic ███████████ ~400
Urdu ██████ ~230
Persian ███ ~130
+ پښتو █ ~40
- dedicated open LLMs for Pashto ......................... 0
+ dedicated open LLMs for Pashto ......................... 1 ← mine
! a language does not die when people stop speaking it.
! it dies when the machines stop listening.So I stopped waiting for someone.
+ scraped, cleaned, deduped a corpus that did not exist
+ fixed tokenization for a script that breaks tokenizers
+ fine-tuned. evaluated. broke it. fixed it. shipped it.
= PAKISTAN'S FIRST PASHTO LLMThe LLM is the loudest thing I've built. It is not the only thing.
|
🤖 AGENTS Systems that plan, call tools, and recover from their own mistakes. LangChain plus custom orchestration. Not chatbots — workers. |
🧬 DATA ENGINEERING Corpora, tokenizers, dedup, quality filtering. The unglamorous layer that decides whether a model is good or garbage. |
👁 VISION Detection, classification, OCR pipelines. Teaching machines to look at something and be right about it. |
|
🎨 GENERATIVE AI RAG, embeddings, vector stores, prompt architecture. Making models produce things actually worth keeping. |
🧠 CLASSICAL ML/DL Before transformers ate the world there was feature engineering, and I still respect it. sklearn, XGBoost, time series. |
🌍 OPEN SOURCE
|
@@ loaded from too many nights @@
+ Python ██████████████████████ 98
+ PyTorch · TensorFlow ████████████████████ 91
+ LLM fine-tuning · LoRA █████████████████████ 94
+ Corpus & tokenizer eng. █████████████████████ 95
+ Agents · LangChain ███████████████████ 85
+ RAG · embeddings · vectors ██████████████████ 82
+ Computer vision · OCR ████████████████ 75
+ Docker · serving · MLOps ███████████████ 71
- Sleep █ 05@@ context: "a language with 40M speakers" @@
+ ▁deserves ████████████████████ 0.91
- ▁lacks ██ 0.06
- ▁waits ▏ 0.02
@@ context: "open models for Pashto:" @@
- ▁zero ████████████████████ 0.97
- ▁several ▏ 0.02
@@ context: "so somebody should" @@
+ ▁build_it ████████████████████ 0.99
- ▁complain ▏ 0.01
- ▁wait ▏ 0.00- refactors working code because the structure "felt wrong"
- 47 tabs open. reads 4. closes none. this is a system, not a flaw.
- allergic to templates. this file is evidence.
- explains a project for 40 minutes when asked "what do you do"
! resists all attempts at normalizationLinkedIn · Email · HuggingFace
Whoever keeps a language alive is writing history.