Raju Sah, Best Java Dev in KTM | Full Stack Java Developer

An AI model trained with reinforcement learning to solve grade-school math problems, explain its reasoning, and produce correctly formatted numerical answers.

Problem solved: Small language models can solve math problems but often produce inconsistent reasoning, malformed answers, or incorrect final numbers. Improving their reliability requires more than asking them to think step by step; the model needs to learn from feedback on its own attempts.

Solution: Built a reproducible reinforcement-learning pipeline using Group Relative Policy Optimization (GRPO), inspired by DeepSeek-R1. Fine-tuned Qwen2.5-1.5B-Instruct on the GSM8K math dataset using Unsloth, TRL, vLLM, and LoRA, with custom rewards for answer correctness, structured reasoning, and valid integer outputs. Trained for 500 steps on a free Kaggle T4 GPU and published the resulting adapter with an interactive demo.

Evaluation: Created a deterministic benchmark using 300 held-out GSM8K test questions, comparing the original model against the GRPO-trained model under greedy decoding. Measured strict answer accuracy, flexible accuracy, output-format compliance, integer compliance, and response length. Also tracked reward progression, training stability, and output quality throughout training.

Result: With the same reasoning prompt, strict answer accuracy improved from 27.00% to 36.33%, an absolute gain of 9.33 percentage points (34.5% relative improvement). Format compliance remained at 99.33%, while integer compliance increased from 42.33% to 83.00%. The experiment demonstrates how reinforcement learning can improve a small model's mathematical reasoning and answer formatting within a limited GPU budget.

A tool that learns the traditional Mithila art style and generates new artwork from text prompts using a custom-trained AI model.

Problem solved: Adapting SDXL to a custom subject or visual style from a small image set can be expensive, difficult to reproduce, and easy to overfit. A usable workflow needs to choose a checkpoint that learns the intended style without simply memorizing training images or losing prompt control.

Solution: Built a reproducible SDXL LoRA training pipeline for 15–40 images on a free Kaggle T4. It prepares the dataset through crop, resize, deduplication, and captioning; trains with Diffusers DreamBooth-LoRA; renders fixed checkpoint-by-prompt comparison grids; and publishes the selected model through a Gradio demo.

Evaluation: Compared Base SDXL with checkpoints from 250 to 1,500 steps across six fixed prompts. Used CLIP for prompt adherence, DINOv2 style similarity against the Mithila dataset, and DINOv2 nearest-neighbor similarity to detect memorization, with a threshold below 0.88.

Result: Checkpoint 750 was selected: its DINOv2 style similarity reached 0.3949, a 111% improvement over the 0.1865 base-model baseline, while retaining a 0.2666 CLIP content score and a 0.6317 Max-NN score. Later checkpoints showed motif repetition and oversaturation, so this checkpoint gave the best balance for deployment.

An AI tool that answers questions using documents and shows where its answers came from. It also tests how accurate and trustworthy those answers are.

Problem solved: Legal research assistants need to give people answers they can verify. A system can retrieve the wrong statute, lose section context, or mishandle Nepali queries; a fluent answer without source-level evidence is unsafe.

Solution: Built a hybrid RAG system for Nepal law that combines structure-aware chunking, BM25 and dense retrieval, reciprocal-rank fusion, cross-encoder reranking, and sentence-level [n] citations. It normalizes Devanagari numerals and routes Nepali queries through a legal glossary so cross-lingual searches retain their meaning.

Evaluation: Added a reproducible harness for Recall@5, MRR, nDCG@5, faithfulness, citation precision/recall, refusal rate, and latency, then compared retrieval, chunking, reranking, and prompting ablations on 1,712 chunks from 1,389 Nepal-law pages.

Result: The full pipeline reached 98.57% Recall@5, 0.9242 faithfulness, and 0.9804 / 0.9254 citation precision / recall. Structure-aware chunking and strict per-sentence citations delivered the largest gains in reliable, traceable answers.

AccountsNepal is a production web application for accounting, billing and inventory management, designed for businesses working across companies, branches and fiscal years in Nepal.

My contribution

I contributed across the Angular frontend and Spring Boot backend, and personally handled and obtained IRD verification for the production system.

Core workflows

• Accounting: double-entry bookkeeping, chart of accounts, journal entries, opening balances, customer receivables and supplier payables.
• Billing and POS: sales invoices, purchase bills, quotations, orders, credit and debit notes, receipts and payment vouchers.
• Inventory and delivery: products, batches, stock movements, stock valuation, split and merge operations, delivery chalans and branch transfers.
• Cash and banking: bank and wallet transactions, fund transfers, cheque tracking and reconciliation workflows.
• HR and payroll: employee records, attendance, leave, salary processing and payroll posting.
• Reporting: general ledger, trial balance, profit and loss, balance sheet, cash flow, stock reports and VAT sales and purchase registers.
• Nepal-specific workflows: Bikram Sambat and Gregorian date support, fiscal-year management, VAT handling and IRD/CBMS synchronization for sales bills and credit notes.
• Administration and exports: company and branch access, role-based permissions, audit logs, data migration, invoice printing and PDF/Excel exports.

🥔 Precision Foliar Pathology & Deep Learning Architecture Suite

Multi-Architecture Neural Ensemble · Explainable AI (Grad-CAM) · Soft-Voting + Temperature Scaling T=0.7

End-to-end foliar pathology suite for clinical diagnosis of potato leaf diseases (Alternaria solani — Early Blight, Phytophthora infestans — Late Blight, and Healthy). Moves beyond single-model prototypes with a Tri-Model Soft-Voting Ensemble (DenseNet-121 w0.45, ConvNeXt-Tiny w0.35, EfficientNetV2-S w0.20, 256×256) and Grad-CAM for transparent, field-ready decisions.

Research Benchmark & Ensemble

• DenseNet-121 / ConvNeXt-Tiny / EfficientNetV2-S — ImageNet-pretrained, fine-tuned, soft-voting + temperature scaling
• Kaggle GPU training, TFLite Float16 (21 MB) & Dynamic INT8 exports for edge/mobile
• ITT: TTA, label smoothing, class-weighted, EarlyStopping

Explainable AI

Grad-CAM saliency maps highlight lesion regions driving each prediction — essential for agronomist trust.

Results

Ensemble outperforms single backbones on 405-image held-out test via soft-voting + calibration. Prior B3 baseline: Test Acc 95.80% / Macro-F1 95.86. See Kaggle notebook for full ablation.

NeuroVision is a real-time face recognition system with live enrollment and instant multi-face identification. Users enroll by capturing guided multi-angle samples from their camera — no offline training, no static dataset required. The system learns new faces on the fly via face embeddings and identifies every person in the live camera feed simultaneously.

• Public Page ui design as well as backend apis.
• In distributor module, referral note, ticket, direct sales, promotional sales , month sales, area sales etc reports.
• In Admin module, distributor reports, sales center reports.
• Inventory management, User management features.

Detects cyberbullying in Formspring Q&A posts using NLP and machine learning. Text is preprocessed (stopword removal, stemming) and vectorized with TF-IDF, then classified using multiple models with hyperparameter tuning via GridSearchCV.

Dataset

Formspring.csv -- 13,147 Q&A text entries from the Formspring social platform, labeled for cyberbullying presence.