🤖 AI/ML & research 44 verified entries

Models, benchmarks and research code from Bangladeshi labs and companies — from BUET's encoder models to Dhaka-built LLM families, plus two MCP servers that bring Bangladesh data into LLM workflows. Each entry notes its licence honestly: community-licence LLMs and non-commercial research checkpoints are labelled, so commercial users know what they can ship before they build on it. Government AI assets are here too, through the banglagov Hugging Face org — the state's first openly published model family.

Edit this section on GitHub · Last verified 2026-11-10 · Rules: curation policy

All entries

• banglagov (Hugging Face) Official Hugging Face org of the Bangla.gov.bd government initiative: banGPT2-Base, banBERT-Base, banELECTRA-Base and banT5-Base pretrained models plus the Ban-Sign-Sent-9K sign-language dataset. The first time a Bangladeshi state body has published model weights openly. Licence terms per model card BD org / for-BD
• BanglaBERT Pretrained Bangla language models from the BUET NLP group — ELECTRA-based encoders that became the default starting checkpoint for Bangla downstream tasks (classification, NER, retrieval) in research and production alike. ⚠ CC BY-NC-SA 4.0 — non-commercial BD org / for-BD
• BanglaNLG BanglaT5 generation and summarisation models from BUET — the natural continuation from their encoder work into text generation for Bangla. ⚠ No licence file BD org / for-BD
• mT5_multilingual_XLSum Multilingual summarisation checkpoint covering Bangla — one of the most widely used ready-made summarizers for Bangla news text on Hugging Face. ⚠ CC BY-NC-SA 4.0 — non-commercial ee list
• normalizer Bangla text normalisation utilities used across the csebuetnlp stack — punctuation, Unicode and encoding cleanup that every Bangla pipeline needs before the first model call. ⚠ No licence file BD org / for-BD
• xl-sum Research code for the XL-Sum summarisation dataset — the 45-language BBC-sourced benchmark that became a standard evaluation for low-resource summarisation, including Bangla. ⚠ No licence file BD org / for-BD
• CrossSum Cross-lingual summarisation across 45 languages — generate a summary in one language from a document in another, built on the same BBC corpus lineage as XL-Sum. ⚠ No licence file BD org / for-BD
• IllusionVQA Optical-illusion visual question answering benchmark — a genuinely adversarial test of whether vision-language models actually perceive or just pattern-match, from the BUET group. ⚠ No licence file BD org / for-BD
• bangla-bert-base Community Bangla BERT — one of the most-used Bangla encoders on Hugging Face, and a permissively licensed alternative to the research-licensed BUET models. MIT ee list
• bangla-bert Training and fine-tuning code for Bangla BERT models — the how-it-was-done repository behind the community encoder above. MIT ee list
• bbocr Bangla handwriting OCR research from the Bengali.AI Kaggle competition lineage — the code behind the grapheme recognition effort that pulled hundreds of Bangladeshi students into ML. BSD-3-Clause BD org / for-BD
• graphemePrepare Grapheme recognition data preparation — tooling for the famous handwritten-grapheme dataset that remains the reference point for Bangla OCR work. MIT BD org / for-BD
• BanglaConformer Conformer-based Bangla speech recognition model — a modern ASR architecture trained for Bangla and published with permissive terms. MIT ee list
• BangalASR Automatic Bangla speech recognition experiments from a SUST researcher — the DeepSpeech-era lineage of Bangladeshi open-source ASR. MIT BD-based
• Bangla-Word2Vec Pretrained Bangla word embeddings — the classic static-embedding layer for Bangla NLP before transformers took over, still useful for lightweight systems. MIT BD-based
• banglanlptoolkit All-in-one Bangla NLP toolkit from the ACI MIS AI and Intelsense AI teams — stemming, POS tagging, NER and more behind one PyPI install. ⚠ Licence unspecified Diaspora
• BanglaBBQ Bias benchmark for question answering in the Bangladeshi context — measuring whether models reproduce social biases when asked about Bangladesh. MIT ee list
• bangla-speech-to-text-benchmark Open benchmark of eight Bengali speech-to-text APIs (Sarvam, Deepgram, Gemini, OpenAI, Groq Whisper and others) evaluated on 1,001 Bangla audio clips — the honest scorecard before you pick an ASR vendor. MIT BD-based
• TituLM Open-weight Bangla LLM family from Hishab (Dhaka), trained on their own 24.3M-document titulm-bangla-corpus; paper arXiv 2502.11187. The most complete locally-published LLM stack: corpus, base models and paper together. Llama 3.2 community licence BD-based
• BanglaLLama Instruction-tuned Bangla LLM from the Bangladesh-led BanglaLLM org — an earlier open-weight 7B entry in the local LLM lineage. Llama 2 community licence BD-based
• Qwen-MediCare-BD Offline-first Bangla medical-assistant model (GGUF builds for local use) fine-tuned on Bangladeshi medicines data — healthcare access where connectivity is weak MIT BD-based
• bangladeshi-bangla-tts-vits Bangladeshi-dialect Bangla text-to-speech built on VITS, with running demo spaces, by a Dhaka-based ML engineer — TTS tuned to how Bangla is actually spoken in Bangladesh rather than Kolkata-standard prose. Apache-2.0 BD-based
• BanglaTag Bangla part-of-speech tagging model by a Dhaka developer — a drop-in POS tagger for pipelines that need morphological context. Apache-2.0 BD-based
• bengali-whisper-medium Whisper fine-tune for Bangla speech recognition from a Dhaka-based maintainer — the practical route to good Bangla transcription on consumer hardware. Apache-2.0 BD-based
• BengaliRegionalASR (Barishal & Sylhet) Dialect-aware ASR trained on BengaliAI's regional-dialect data for two Bangladeshi divisions — addressing the reality that Barishai and Sylheti Bangla break standard ASR models. Apache-2.0 BD org / for-BD
• @entrogic-net/mcp-server-bangladesh MCP server exposing Bangladesh reference data (geography, utilities and more) to LLM clients, from the BanglaClaw team — so AI assistants can answer BD questions with real data instead of guessing. Apache-2.0 BD-based
• bangla-news-mcp MCP server that retrieves live Bangla news into LLM contexts, by Abid Mohammad Jawad (open source, GitHub hurutta/bangla-news-mcp) — current-events grounding for Bangla-speaking AI workflows. see notes BD-based
• SellerMate BD Self-hosted AI assistant (Ollama) that writes Bangla/English product descriptions and SEO copy for BD sellers — generates Bangla/English product descriptions, key features, SEO metadata, Facebook captions, and FAQs using a locally hosted Ollama model. MIT BD-based
• DocTrust Reviewer agent that cross-examines BD identity document sets (NID, birth certificate) and flags mismatches — reads every field of every document in Bangla, English and Hindi ... deterministic code decides the verdict; measured on 40 synthetic Bangla sets. MIT BD-based
• Gochalo (গোছালো) Bilingual AI workspace that turns Bangladeshi notices and circulars into tasks, deadlines and document plans — starting audience is Bangladeshi students and families navigating everyday applications and notices; deployed with FastAPI backend. MIT BD-based
• ClinicDesk AI Bilingual AI front desk for small clinics: answers patients on WhatsApp and web, books, escalates to staff — most small clinics in Dhaka get their patient enquiries on WhatsApp ... books appointments, and passes the conversation to a staff member when it should. MIT BD-based
• MedEdge AI Pipeline that turns paper lab reports and doctor dictation into verified HL7 FHIR bundles, anonymising pages before any model reads them. Apache-2.0 BD org / for-BD
• BanglaHealth-AI Explainable diabetes and heart risk stratification for low-resource BD clinics: SHAP-backed three-tier scores, FastAPI, Streamlit dashboard. MIT BD-based
• Bangla Sign Language Detector Reproducible Bangla hand-sign recognition (CNN vs MobileNetV2) with image, video and live-camera inputs plus a PyQt5 app; tested on KU-BdSL. MIT BD org / for-BD
• LAI — Local AI & Android Automation Runtime Source-only Android runtime for private on-device LLMs (llama.cpp/GGUF) with a Bangla-first interface and consent-gated device automation. Apache-2.0 BD-based
• BdSL Vision Bangladesh Sign Language recognition pipeline: 4,139 annotated samples, 18 classes, YOLOv8s training and ONNX browser-inference demo — "4,139 annotated samples · 18 classes · YOLOv8s · mAP@0.50 0.639... Real ONNX inference runs locally in the browser with image upload and camera input, Bengali labels.". ⚠ No licence file BD-based
• SignBridge Experimental web app pairing Bangla videos with saved avatar-signing tracks (BdSL), clearly marking which meaning it could not sign. ⚠ No licence file BD-based
• Bengali Books Engine Multi-agent engine that plans, writes, reviews and publishes complete Bengali books from a single CLI; ships 7,800+ generated chapter files. ⚠ No licence file BD org / for-BD
• BanglaT5 (BUET NLP) Canonical Bangla T5 seq2seq model from the BUET NLP group (summarisation, NLI, NER, QA). MIT (per model card) BD org / for-BD
• Bangla Twoclass Sentiment Analyzer Most-downloaded Bangla sentiment model on Hugging Face (3.3M+ downloads), positive/negative classification. MIT BD org / for-BD
• wav2vec2 XLS-R 300M Bengali Fine-tuned XLS-R 300M automatic speech recognition model for Bengali (670k+ downloads). Apache-2.0 BD org / for-BD
• BanglaASR (bangla-speech-processing) Community Bangla automatic speech recognition model maintained by the bangla-speech-processing org. MIT BD org / for-BD
• Bengali.AI Whisper-medium ASR Whisper-medium fine-tune for Bangla speech recognition from the Bengali.AI community. Apache-2.0 BD org / for-BD
• BanglaGPT-Pro Bangla text-generation model (Gemma-based) for local inference. Gemma (custom) BD org / for-BD

How this section is verified

• Every link is fetched live at verification time — dead links are removed, not kept for history.
• Bangladesh connections are evidenced (profile location, org page, or the product's own docs) — see the scope rules.
• Dormant items (3+ years without activity) move to the legacy file with their last-activity date.
• Licence information is quoted as the publisher states it; gaps are flagged with ⚠️ rather than guessed.