Two systems show full arc from missing resource or brittle toolchain to result others can inspect and extend.
R01 · Retrieval system · Paper + code + Space
Winner, aIDEA National AI Hackathon 2026
NepSAUL: Nepali legal question answering
Legal retrieval study over 10,320 Nepali Supreme Court case laws, comparing sparse and multilingual dense methods before grounding answer generation in retrieved evidence.
BM25 chunks91%Dense retrieval75%
Research question
Which retrieval strategy provides most reliable first result for Nepali legal queries?
Compared
BM25 chunk and document retrieval against multilingual-e5-large, e5-base, and LaBSE dense retrieval with FAISS.
Evaluation
Precision@1 for retrieval, followed by LLM-as-Judge and legal-expert review for generated answers.
Finding
BM25 chunks reached 91% Precision@1. Reported dense setup reached 75% on same task.
Next work
Agentic multi-step legal reasoning and document synthesis remain in progress, not reported as completed results.
M01 · Speech model · Checkpoint + app code
Public checkpoint and app code
Chatterbox Nepali: adapting speech toolchain
Useful contribution was not wrapper around pretrained model. It was making training and inference path handle Nepali text, reference audio, and hardware available for iteration.
नेपाली inputDevanagariModel500M TTS
Adaptation
Fine-tuned Chatterbox Multilingual 500M for Nepali speech.
Tooling
Patched tokenizer and alignment analysis for Devanagari character clusters.
Access
Released checkpoint, Gradio app code, and Apple Silicon training support.
Open question
Formal quality benchmark plus explicit reference-voice consent and data-provenance notes remain needed. Release is inspectable model artifact, not finished benchmark claim.
02
Questions behind systems.
Each thread starts with practical gap, then moves toward benchmark, model, or public artifact.
Q01
Speech and voice
What breaks when tokenizer and alignment stack were not designed for Nepali?
Adapt whole path, not only checkpoint. Devanagari-aware tokenizer and alignment changes, Apple Silicon training, public checkpoint, and companion app code make result inspectable.
Q02
Retrieval and grounding
What does lexical retrieval reveal when multilingual embeddings underperform?
Benchmark simple baseline before adding complexity. On NepSAUL corpus, sparse BM25 reached 91% Precision@1 while reported dense setup reached 75%.
Q03
Language data
Which annotation and evaluation choices make low-resource benchmark useful?
When benchmark is missing, build evidence base. Nepali offensiveness work began with collection, guidelines, cleaning, and quality control for 11.6K samples.
03
Models, data, code, and serving paths.
Inspectable today. Evidence and status stay separate from ambition.
Type
Artifact
Contribution
Evidence
Access
R01 · Retrieval
NepSAUL research frameworkNepali legal QA
Corpus pipeline, sparse and dense retrieval, generation evaluation
Five distinct studies. Methods, evaluation, findings, limits, and publication status stay visible; duplicate indexes and preprint entries are not counted twice.
P05 · Multimodal NLP · Accepted paper + preprint
Accepted · CHiPSAL at LREC 2026
MEME-Fusion: what visual context adds to Nepali meme understanding
Controlled ablation study of hate detection and sentiment analysis on Nepali memes, built to measure whether image context helps when text and vision resources are both limited.
Samir Wagle, Reewaj Khanal, Abiral Adhikari Second Workshop on Challenges in Processing South Asian Languages, co-located with LREC 2026. · 2026
Fusion gain+5.9 F1Configurations8
Research question
Does explicit cross-modal fusion improve Nepali meme classification beyond text-only and image-only baselines?
Compared
Eight configurations using BGE-M3 text features, CLIP ViT-B/32 visual features, self-attention, learned gating, late fusion, and ensemble variants.
Evaluation
Macro F1 across hate-detection and sentiment tasks, with component ablations used to isolate where gains came from.
Finding
Explicit cross-modal fusion improved macro F1 by 5.9 points over text-only baseline on hate detection.
Limitation
English-centric visual encoder handled Devanagari text inside images poorly; ensembles were unstable with roughly 850 samples per fold.
P04 · Legal NLP + RAG · IEEE conference paper
Published · ICNLP 2026
NepSAUL: retrieval-augmented question answering for Nepali law
Retrieval and grounded-answer study over 10,320 Nepali Supreme Court case laws, designed around language, legal vocabulary, and evidence constraints missing from general-purpose RAG benchmarks.
Samir Wagle, Abiral Adhikari, Reewaj Khanal, Batsal Bhandari 8th International Conference on Natural Language Processing, ICNLP 2026. · 2026
BM25 chunks91%Dense retrieval75%
Research question
Which retrieval strategy returns most reliable first result for Nepali legal questions?
Compared
BM25 chunk and document retrieval against multilingual-e5-large, e5-base, and LaBSE dense retrieval with FAISS.
Evaluation
Precision@1 for retrieval; LLM-as-Judge and legal-expert review for generated-answer truthfulness and grounding.
Finding
BM25 chunk retrieval reached 91% Precision@1; reported dense configuration reached 75% on same task.
Next work
Agentic multi-step legal reasoning and document synthesis remain active work, not reported as completed results.
P03 · Representation study · Journal paper
Published · KU JSET
Which sentence embeddings work for Nepali sentiment?
Factorial comparison of multilingual sentence representations and downstream classifiers, asking whether stronger embeddings matter more than adding classifier complexity for low-resource sentiment analysis.
Abiral Adhikari, Samir Wagle, Reewaj Khanal, Prashant Manandhar Kathmandu University Journal of Science, Engineering and Technology, Vol. 18, No. 2. · 2024
Best accuracy82.49%Model pairings12
Research question
For Nepali sentiment classification, how much comes from embedding quality and how much from downstream model complexity?
Compared
BGE-M3, LaBSE, multilingual-e5-base, and DistilUSE embeddings across MLP, residual MLP, and Transformer classifiers.
Evaluation
Twelve embedding-classifier pairings evaluated on same Nepali sentiment task using accuracy and class-level results.
Finding
BGE-M3 with simple MLP reached 82.49% accuracy and beat more complex downstream alternatives in this evaluation.
Implication
Representation quality dominated classifier complexity; stronger encoder plus simpler head was best measured trade-off.
P02 · Dataset + classification · ICON paper
Published · ICON 2024
Separating profanity from offensiveness in Nepali text
Dataset and classification study treating profanity and offensiveness as related but different problems, instead of collapsing both into one moderation label.
Abiral Adhikari, Prashant Manandhar, Reewaj Khanal, Samir Wagle, Praveen Acharya, Bal Krishna Bal Proceedings of 21st International Conference on Natural Language Processing, pages 515–521. · 2024
Profanity87.8%Offensiveness78.2%
Research question
Can Nepali moderation models distinguish explicit profanity from broader offensive meaning?
Data
11,681 Nepali social-media comments collected, cleaned, and manually annotated for task-specific labels.
Compared
BiLSTM classifiers, vocabulary embeddings, and multilingual transformer representations across binary and three-class settings.
Evaluation
Task accuracy for binary profanity, binary offensiveness, and multi-class classification.
Finding
BiLSTM reached 87.8% on binary profanity and 78.2% on binary offensiveness; broader multi-class task remained harder.
P01 · Digital media ethics · Public study
Experiment findings · not peer reviewed
What clickbait costs Nepal’s digital information ecosystem
Mixed-method study of ethical tension between short-term engagement and long-term trust across Nepal’s blogs, YouTube channels, and e-commerce platforms.
Reewaj Khanal, Abiral Adhikari, Samir Wagle, Hempal Shrestha Public research record, June 2024. · 2024
MethodMixedContexts3
Research question
How do clickbait practices affect trust, misinformation, and responsibility across Nepal’s digital media ecosystem?
Scope
Blogs, YouTube, and e-commerce—three contexts where attention incentives shape headline and thumbnail choices.
Method
Survey and interview evidence interpreted through utilitarian, fairness, virtue-ethics, and common-good frameworks.
Finding
Clickbait may create short-term engagement but compounds misinformation risk and long-term trust damage.
Status
Public experiment findings, not counted as peer-reviewed publication. Record kept visible so evidence status stays explicit.
05
Research depth with systems range.
Ownership across data, model, evaluation, API, and surrounding system constraints.
Freelance AI Engineer
Voice and agentic systems
ASR-to-LLM-to-TTS agents, retrieval systems, telephony and CRM integrations, automation workflows, and deployment infrastructure.
Speech AI Engineer
Mantra Ideas
Nepali and multilingual ASR/TTS adaptation; speech data collection, cleaning, normalization, forced alignment, augmentation, and low-latency inference.
R&D Engineer
CityTech
Embedded C and Android software for payment terminals: TCP/MQTT, Android 10-to-14 migration, NFC, card-reader, and PIN-pad integration.
Undergraduate Researcher
KU Information & Language Processing Lab
Nepali sentiment and offensiveness models, datasets, multilingual embeddings, and English–Nepali–Tamang machine translation through Google’s Society-Centered AI Research Program.
06
How I approach under-resourced problems.
Baseline first. Missing data visible. Whole toolchain inspected. Public artifact left behind.
01
Start with baseline.
NepSAUL kept BM25 and found it ahead of dense retrieval for measured task.
02
Build missing data.
Offensiveness work treated collection, annotation design, and cleaning as research contributions.
03
Inspect whole stack.
Speech adaptation included tokenizer, alignment, hardware, and serving constraints.
04
Leave artifact.
Code, datasets, model cards, demos, and limitations make work verifiable and extensible.
Research collaboration
Let’s investigate languages current stack misses.
Interested in multilingual speech, low-resource evaluation, retrieval, and models that must survive outside benchmark.