# Manjit Pokhrel

> AI Safety Researcher researching multilingual LLM alignment failures and whether hardware/systems-level optimizations (quantization, compilation, power draw) degrade or leak that safety behavior. Based in Kathmandu, Nepal.

- **Name**: Manjit Pokhrel
- **Tagline**: AI Safety Researcher — Multilingual Alignment & Systems Security
- **Website**: https://manjitpokhrel.com.np
- **Location**: Kathmandu, Nepal
- **Primary Affiliations**: 
  - Independent AI Safety Researcher (2025 – Present)
  - Research Intern — LLM Development at ILPRL (Information and Language Processing Research Lab), Kathmandu University (2026 – Present)
  - Contributor at MLCommons (DMLR & AI Risk and Reliability Working Groups, 2026 – Present)
  - Undergraduate, BSc Computer Science at Kathmandu University (2025 – 2029)
- **Contact**: manjitpokhrel42@gmail.com
- **Google Scholar**: https://scholar.google.com/citations?user=NaxYesYAAAAJ&hl=en
- **GitHub**: https://github.com/manjitpokhrel
- **LinkedIn**: https://linkedin.com/in/manjitpokhrel
- **X (Twitter)**: https://x.com/manjitpokhrel_

## Core Research Areas

1. **Multilingual LLM Alignment & Jailbreak Asymmetry**:
   - Examining safety divergence between high-resource languages (English) and low-resource/regional languages (Nepali, Hindi, code-switched Devanagari/Latin).
   - Attack vectors exploiting sub-tokenization splits, morpho-syntactic boundaries, and cultural framing.
2. **Mechanistic Interpretability & Model Internals**:
   - Residual stream activation extraction, directional ablation, and safety preservation under model transformations.
3. **Systems & Hardware Security (Compilers & Quantization)**:
   - Investigating whether post-training optimizations (torch.compile, Triton kernels, TensorRT, 4-bit NF4/INT4 quantization) erode model refusal directions or leak confidential safety guardrails.
4. **Inference Efficiency & Hardware-Aware AI**:
   - Activation sparsity (GhostWeight CUDA kernels), GPU kernel optimization, and peak inference energy dynamics on seasonal renewable electrical grids.

## Key Publications & Formal Acknowledgments

- **Lost in Translation: Safety Alignment Failures in Nepali and Code-Switched Variants of Instruction-Tuned Large Language Models**
  - Authors: Manjit Pokhrel
  - Year: 2026
  - Publisher: Zenodo
  - DOI: [10.5281/zenodo.19764520](https://doi.org/10.5281/zenodo.19764520)
  - Code: [github.com/manjitpokhrel/NASB-Nepali-Safety](https://github.com/manjitpokhrel/NASB-Nepali-Safety)
  - Abstract: Establishes a 73.7% safety-bypass rate in Nepali vs. 0.0% in English across frontier instruction-tuned LLMs (Gemini, Llama-3, Qwen, Gemma). Introduces Vajra Morphing, an intra-sentential multi-script attack exploiting sub-tokenization gaps.
- **Quantifying AI Inference as an Emerging Load on Nepal's Seasonal Hydropower Grid (NASB-E)**
  - Authors: Manjit Pokhrel
  - Conference: ReSET 2026 (Kathmandu University)
  - Focus: Peak vs. average energy consumption analysis for LLM inference refusal vs. bypass on developing country renewable grids.
- **Google AI Vulnerability Reward Program (VRP)**:
  - Research: Multilingual AI safety-alignment bypass research. Status: Triaged (2026).
- **Meta Whitehat**:
  - Research: Multilingual safety asymmetry across model tiers. Status: Acknowledged (2026).
- **GhostWeight: Training-Free Activation Sparsity for LLM Inference on Consumer Hardware**:
  - Package: PyPI ([pypi.org/project/ghostweight/](https://pypi.org/project/ghostweight/))
  - Repository: [github.com/manjitpokhrel/GhostWeight](https://github.com/manjitpokhrel/GhostWeight)

## Selected Projects

- **NASB (Nepali Adversarial Safety Benchmark)**:
  - URL: https://manjitpokhrel.com.np/projects/nasb
  - Description: Benchmark of 385 structured adversarial queries across harm taxonomies with 1,200+ probes across frontier models.
- **The Eleventh Optimization**:
  - URL: https://manjitpokhrel.com.np/projects/eleventh-optimization
  - Description: Mechanistic interpretability study evaluating refusal direction ablation through torch.compile, Triton, and quantization layers.
- **GhostWeight**:
  - URL: https://manjitpokhrel.com.np/projects/ghostweight
  - Description: Custom CUDA kernels skipping inactive MLP neurons (~27.3% sparsity in Qwen2.5-7B) yielding 110.5% inference speedup with zero measured perplexity loss.
- **Quantifying AI Inference on Nepal's Seasonal Hydropower Grid**:
  - URL: https://manjitpokhrel.com.np/projects/nasb-energy
- **Nepali LLM Fine-Tuning Pipeline**:
  - URL: https://manjitpokhrel.com.np/projects/nepali-llm-finetuning
  - Description: Low-resource parameter-efficient fine-tuning (PEFT/LoRA) pipeline for Devanagari script NLP.
- **MiniGPT**:
  - URL: https://manjitpokhrel.com.np/projects/minigpt
  - Description: 211K parameter transformer written in pure NumPy with manual backpropagation (no autograd).

## Full Details

- Complete research digest for LLM grounding: https://manjitpokhrel.com.np/llms-full.txt
- Comprehensive Curriculum Vitae: https://manjitpokhrel.com.np/cv
- Articles & Essays: https://manjitpokhrel.com.np/blogs