Research & Insights

In-depth empirical research, technical benchmarks, and system evaluations across low-resource NLP, ASR, and autonomous agents.

Benchmarking & LLMsPDF

Himalayan GPT for Restaurant Ordering Chatbots: Benchmarking Intent Classification, High Inference Latency, and Free Generation in Low-Resource Nepali

An empirical benchmark of Himalayan GPT models for Nepali restaurant ordering chatbots. We evaluate intent classification accuracy (92.31%) with himalaya-gemma-4-e2b-it, uncover severe 39.6s inference latency due to CPU/disk offloading on laptop GPUs, and diagnose generation collapse in himalayagpt-0.5b.

Dimanjan Dahal, Kushal Shrestha, Gemini Antigravity
14 min read
7/22/2026
Model Engineering & Fine-TuningPDF

Technical Investigation: Why Fine-Tuning HimalayanGPT 0.5B Fails on the Hugging Face Stack (and Why Qwen2.5-3B Succeeded)

A root-cause technical investigation into why fine-tuning himalaya-ai/himalayagpt-0.5b with QLoRA failed at step 0 with a CUDA device-side assert. We contrast its custom Nanochat architecture (8 value_embeds tables, invalid special token ID 32768, and ignore_index=-1) with Qwen2.5-3B (which achieved 99.73% accuracy on the same 14,000 Roman Nepali dataset).

Dimanjan Dahal, Kushal Shrestha, Gemini Antigravity
18 min read
7/27/2026
Speech & Audio AIPDF

Benchmarking Nepali Automatic Speech Recognition: An Empirical Evaluation of Kriti vs. NepaliConformer Offline on Real-World Speech

A comprehensive empirical evaluation of two leading Nepali ASR models—Kriti (harrrshall/kriti) vs NepaliConformer Offline (ampixa/nepali-conformer-offline)—on the shunyalabs/nepali-speech-dataset. Kriti achieves a 31.19% WER and requires only 0.74 GB VRAM, while NepaliConformer delivers 219.7x real-time speed.

Dimanjan Dahal, Kushal Shrestha, Gemini Antigravity
15 min read
8/20/2026
Computer Vision & RetrievalPDF

Furniture Matching Engine Evaluation: Benchmarking Geometric and Keypoint Retrieval on Held-Out Product Imagery

An empirical evaluation of a multi-signal visual retrieval engine (LSD line geometry, SIFT/affine keypoints, material signature) on 26 held-out furniture images against a 161-item catalogue. While top-5 retrieval reaches 69.2%, top-1 category accuracy is 38.5% and a 45% confidence threshold creates a 100% false-pass rate.

Dimanjan Dahal, Kushal Shrestha, Gemini Antigravity
13 min read
9/9/2026
Case Study

Improving E-Commerce Search: A Case Study in Hybrid Retrieval

A detailed engineering case study on how we achieved a 15% improvement in Mean Reciprocal Rank (MRR) by implementing a weighted hybrid search system combining keyword matching with neural semantic understanding.

Sports Center Nepal Engineering Team
12 min read
1/21/2026
Research

Modern Enterprise AI: Dynamic Adaptation, Governance & Agentic Autonomy

Comprehensive analysis of modern enterprise AI imperatives covering dynamic adaptation, robust governance, agentic autonomy, build vs. buy decisions, and strategic recommendations for C-suite leaders achieving competitive advantage through ethical AI implementation.

Sajedar Research Team
19 min read
1/25/2025
Research

The AI Imperative in CX: Advanced Conversational AI Framework

Comprehensive framework for implementing advanced conversational AI in customer experience, covering GenAI architecture, multimodal capabilities, brand voice consistency, smart escalation, and 24/7 CX revolution achieving up to 8x ROI.

Sajedar Research Team
22 min read
1/25/2025
Research

AI Agents in 2025: Strategic Imperative & Architecture of Intelligence

Comprehensive analysis of AI agents as strategic business imperatives in 2025, covering market dynamics, RAG architecture, human-AI convergence, ROI measurement, and governance frameworks for enterprise success.

Sajedar Research Team
20 min read
1/25/2025
Research

Conversational AI: The Strategic Imperative for Autonomous Sales

Comprehensive analysis of conversational AI as a strategic business imperative, covering 24/7 lead qualification, intelligent profiling, conversion acceleration, and technical blueprints for autonomous sales systems achieving up to 8x ROI.

Sajedar Research Team
18 min read
1/25/2025
Research

Catalyzing National Progress: 40 AI Chatbot Projects For Nepal's Budding Innovators

A comprehensive guide to 40 AI agent projects across agriculture, tourism, manufacturing, public services, and financial inclusion that can transform Nepal's economy and empower local innovators.

Sajedar Research Team
18 min read
1/20/2025
Research

Nepal's Conversational AI Research Agenda: A Comprehensive Framework for 2025

A formal research agenda to support Nepal's conversational AI ecosystem, featuring 40 concrete research questions across foundational NLP, advanced models, sector-specific applications, and ethical AI development.

Sajedar Research Team
15 min read
1/15/2025
Research

The Conversational AI Frontier: Nepal & South Asia, 2025

Comprehensive analysis of conversational AI adoption, market trends, and strategic opportunities across Nepal and South Asia in 2025.

Sajedar Research Team
12 min read
1/1/2025
Research

The Architecture of Dialogue: Conversational UX Best Practices (2025)

Comprehensive guide to designing human-like conversational experiences. From foundational principles to advanced dialogue mechanics and error handling strategies.

Sajedar Research Team
15 min read
1/1/2025
Research

The Definitive Guide to Chatbot ROI (2025)

Comprehensive guide to calculating and maximizing agent ROI. From strategic imperatives to detailed cost-benefit analysis and implementation frameworks.

Sajedar Research Team
18 min read
1/1/2025
Chat on WhatsApp