Sumit K C

Retrieves governing Medicare NCD/LCD policies (such as ICD, PAP, TAVR) using hybrid lexical-semantic chunk matching, then extracts clinical criteria with character offset anchors.

An adversarial verification agent programmatically confirms character slices directly against the raw EHR text to prevent LLM hallucinations, quote drifts, and negative omissions.

Enforces a calibrated confidence gate (τ = 0.85); cases falling below the threshold route safely to human auditors, while verified decisions generate complete CMS appeal defense packets.

Run equivalent English and Nepali prompts through the model gateway. The Flutter client connects to FastAPI to monitor experiments and inspect paired responses.

Policy retrieval supplies context to a LangGraph Judge/Critic/Arbiter workflow. The agents evaluate responses, challenge classifications, and reconcile disagreements. Adaptive red-teaming explores prompt variations.

Compare refusal, unsafe compliance, and over-refusal behavior using paired statistical tests and bootstrap confidence intervals. Execution traces track latency, token usage, and estimated costs; CLI tools support dataset verification and regression checks.

Add ingredients and dietary preferences in the React pantry interface, then request recipe suggestions or a cooking answer.

The Node.js API connects to the FastAPI RAG service, which retrieves indexed recipes relevant to the pantry or question.

Gemini uses retrieved recipe context to generate suggestions and cooking answers. Ranked retrieval and structured responses provide a fallback when Gemini is unavailable.

Admins manage traffic signs, question banks, exam papers, tutorials, and notices. Scheduled official-source imports receive Gemini screening, with human review required before publication.

Published resources feed the learner website and Flutter library through the Laravel API, with search, language filters, and saved resources.

Learners choose English or Nepali, review traffic-sign flashcards, and start randomized practice or a 30-minute mock exam with up to 25 questions.

Attempts can be saved and resumed. The server grades submissions and reveals answers and explanations after submission, using shared attempt records across web and mobile.

My learning tracks daily goals, streaks, accuracy, milestones, and practice history, and suggests a next step based on unfinished attempts or topics needing attention.

Premium learners can use personalized practice, revision tools, and a Gemini study assistant. Private support conversations and content reports connect learners with admins.

Approved schools manage Owner, Manager, Instructor, and Student memberships through invitations and role-based access. Courses and capacity-limited batches connect students with instructors, scheduled classes, and attendance records.

School staff maintain student exam and licence history, record instructor feedback, and review reports. Students control whether their completed-practice totals are shared with staff.

Private lessons include instructor vehicles, available slots, conflict-checked booking, rescheduling, waitlists, and optional server-verified Khalti/eSewa payments. In-app notifications and scheduled reminders keep members informed.

The Flutter school centre shares the Laravel services; school management and booking workflows also use the responsive website through secure browser handoffs.

Vision agents analyze the uploaded photo to identify the item that could be reused.

Research agents find sustainable reuse options and DIY recommendations for the identified item.

A verification agent checks the recommendations before the coordinated pipeline presents a reuse plan to the user.

Learners practice real-world dialogues with an AI conversation partner powered by streaming SSE responses from Google Gemini and Gemma2.

Real-time speech evaluation reviews pronunciation, identifies sentence structure mistakes, and offers instant corrections.

Mongoose-backed progress controllers record practice streaks, mastery milestones, vocabulary retention, and smart scheduling reminders.

The SMS Spam Collection dataset is tokenized, stripped of stopwords, and stemmed, then represented using Bag-of-Words and TF-IDF.

Compare Naive Bayes, Logistic Regression, Linear SVM, Decision Tree, Random Forest, and KNN.

Linear SVM achieved 98.1% accuracy with the best precision/recall balance in the project evaluation and was selected for final deployment.