Kannan

I am a 12th-grade student at DPS International School, Singapore, dividing my time between AI research, competitive informatics, and full-stack software development. My work focuses heavily on analyzing Large Language Model (LLM) limitations, building democratized EdTech solutions, and algorithmic problem-solving.


1. Education & Academics

DPS International School

Singapore | Jul 2016 – Expected Mar 2027

Class 12 ISC (Indian School Certificate)


2. Publications & Talks

In reverse chronological order

From Semantic Reuse to Auditable Tutoring

Research Essay, 2026

This research proposes a "glass-box" extension of the Crowdsourced Semantic Cache Network (CSCN) to address the computational and governance barriers to scaling AI tutoring in education. By introducing a multi-stage pipeline that integrates safety screening, privacy transformation, and intent canonicalization, the framework shifts the unit of semantic reuse from raw conversational text to structured, pedagogically aligned "tutoring nodes." This approach significantly reduces marginal inference costs through semantic caching while ensuring that AI interactions remain transparent, Socratic, and fully auditable for institutional oversight, providing a scalable blueprint for governance-compliant educational AI.

View Essay

Open Source Summit India, 2026

This talk introduces Statutory Myopia — a newly documented failure mode in Retrieval-Augmented Generation (RAG) systems, where LLMs over-rely on retrieved text and suppress their own pre-trained knowledge of conflicting or superseding legal rules.

Using an open evaluation harness built on the LegalBench dataset (3 doctrinal task subsets, 239 unique cases, 4,302 reasoning traces), three frontier models — GPT-5, DeepSeek R1, and Gemini 2.5 Pro — were benchmarked across No-RAG and Statute-Only RAG conditions. The results were counterintuitive: incomplete retrieval caused a 9 percentage point drop in accuracy globally (0.87 → 0.78), with a worst-case 21 pp decline in the Hearsay task when the superseding Crawford v. Washington ruling was withheld from the context.

The talk also covers the custom async evaluation harness with SHA-256 cryptographic caching for reproducibility, the blacklist protocol for controlled retrieval experiments, and implications for any hierarchical domain — including medical guidelines, API deprecation, and compliance specifications.

View Presentation Slides | Watch the Talk

Cracking the Code: From Caesar Ciphers to WhatsApp Encryption

STEAM Club, 2026

This talk traces the evolution of cryptographic thinking from antiquity to the modern smartphone era. Beginning with the Caesar cipher — a shift substitution scheme dating to approximately 100 BC — the session establishes core terminology (plaintext, ciphertext, encryption, decryption) before demonstrating why simple substitution ciphers are fundamentally insecure: with only 26 possible keys and English letter frequency distributions preserved under any shift, such ciphers are trivially broken by brute force or frequency analysis. The talk then addresses the central challenge of modern cryptography — secure key exchange over a public channel — and introduces RSA (Rivest, Shamir & Adleman, 1977) as the solution. The mathematical foundations of RSA are presented through a worked example, grounding the public/private key paradigm in modular arithmetic and the computational hardness of integer factorisation. The final segment applies these principles to contemporary practice: WhatsApp's end-to-end encryption via the Signal Protocol is examined alongside its underlying primitives (Curve25519, AES-256, HMAC-SHA256), and the distinction between cryptographic content protection and metadata privacy is used to motivate a comparison of WhatsApp and Signal as messaging platforms. The talk concludes by situating cryptography within everyday digital life, noting that every HTTPS connection constitutes applied public-key cryptography.

View Presentation Slides

Crowdsourced Semantic Cache Network: A Distributed, User-Funded Knowledge Network for Cost-Efficient and Self-Correcting LLM Inference

Preprint, 2026

Murugapandian, Kannan

The widespread deployment of Large Language Model (LLM) inference APIs presents a fundamental economic challenge: the near-total redundancy of user queries drives operational costs that scale linearly with the number of requests, regardless of semantic overlap between them. This paper proposes and formally analyses a novel four-layer architecture—the Crowdsourced Semantic Cache Network (CSCN)—that addresses this challenge through three complementary mechanisms: (i) a globally shared semantic vector database that intercepts semantically equivalent queries before they reach the inference layer; (ii) a user-triggered, single-token LLM-as-judge validation gate that replaces time-based cache invalidation with demand-driven accuracy verification; and (iii) a freemium token-economics model that converts user payments into a compounding, community-maintained knowledge graph. We formalise the architecture using cosine similarity over high-dimensional embedding spaces, develop closed-form cost models for all system states, and derive an expected daily savings function S(N,H) across the full range of empirically observed cache hit rates. Under conservative production parameters (N = 1,000,000 queries per day, H = 0.40), the CSCN yields a 39.9986% reduction in raw LLM API expenditure; under optimistic but achievable parameters (H = 0.67), savings reach 66.9977% of baseline cost. Validation calls are shown to be approximately 2.8827× cheaper than full generation calls, enabling a gross margin of approximately 27.5% on paid-tier operations. The architecture is further demonstrated to exhibit positive network externalities, wherein marginal inference cost approaches zero as the knowledge base grows.

DOI: 10.5281/zenodo.19401231 | Plain English Breakdown

Preprint, 2026

Murugapandian, Kannan

Retrieval-Augmented Generation (RAG) is widely assumed to improve Large Language Model (LLM) accuracy by grounding outputs in factual sources. We challenge this assumption in legal reasoning contexts. Evaluating GPT-5, DeepSeek R1, and Gemini 2.5 Pro across 239 legal cases drawn from three LegalBench tasks, we find that providing statutory text without accompanying case law degrades ensemble majority-vote performance by 9 percentage points (0.87 → 0.78). We term this phenomenon Statutory Myopia—the tendency of models to privilege explicitly retrieved statutory text over implicit pre-trained legal knowledge, including controlling precedent. The effect is most pronounced in models assigned a Textualist judicial persona (−12.1 pp) and in the Hearsay task, where retrieval withheld a superseding constitutional precedent (−21 pp). Our findings suggest that incomplete retrieval may harm legal reasoning more than no retrieval at all.

DOI: 10.5281/zenodo.19522950 | Plain English Breakdown | Plain English Breakdown of the Technical Setup

Democratizing Personalized Learning using Open Source AI

FOSSASIA Summit '26

Presented at the FOSSASIA 2026 Summit. Explored the intersection of AI and Bloom's 2 Sigma Problem, arguing that reliance on closed-source LLMs introduces severe privacy risks. Proposed a "Glass-Box Learning Architecture" utilizing locally-hosted, fine-tuned 7B open-source models via LoRA, combined with Verified Retrieval (RAG) and secondary guardrail classifiers to deploy cost-effective, Socratic AI.

View Presentation Slides | Watch the talk

Introduction to Digital Forensics

DPS International School, Mar 2026

Designed and delivered a comprehensive cybersecurity presentation to junior students. Broke down the fundamentals of Digital Forensics for Capture The Flag (CTF) competitions, covering Magic Bytes, PCAP network analysis via Wireshark, LSB Steganography, and Disk Imaging using Autopsy.

View Presentation Slides

Artificial Intelligence in Education

Medium, Jun 2025

Authored a research article analyzing inequities in modern education systems, examining 18+ peer-reviewed studies on teacher quality disparities and curriculum biases. Proposed AI-powered personalized learning as a scalable solution.


3. Experience & Leadership

Research Team Lead

Sep 2025 – Nov 2025

The Junior Academy, New York Academy of Sciences

View Presentation Slides

Co-Founder & Co-Head

Dec 2021 – Oct 2024

Earth Status (Youth Climate Action Initiative)

Museum Ambassador

Aug 2025 – Present

HeritageSG, National Heritage Board


4. Honors & Awards


5. Technical Projects

LeginAI | Multi-LLM Reasoning System

Orchestrated a 9-agent consensus workflow using voting logic to minimize hallucinations in legal verdict generation. Implemented a robust Retrieval-Augmented Generation (RAG) pipeline to automatically retrieve and ground agent outputs strictly within uploaded legal statutes.

DonumAI | AI-Powered Teaching Assistant

Developed an educational technology platform addressing teaching inequality by generating personalized lessons tailored to student comprehension levels. Implemented a Flask backend integrating the Google Gemini API with custom JSON-to-Markdown parsers for dynamic, real-time content rendering.

Live Demo

Learn with DonumAI | AI-Curated Knowledge Podcast

A daily podcast series produced in conjunction with the DonumAI platform, designed to make structured, AI-curated knowledge accessible in an audio format. Each episode distills lesser-known scientific phenomena, overlooked historical events, and interdisciplinary insights into concise, research-backed segments. The series targets intellectually curious listeners seeking substantive content beyond mainstream discourse, combining the precision of AI-assisted curation with accessible science communication.

Listen on Podbean

MittentisAI | Narrative Generation Engine

Engineered a storytelling engine utilizing modular prompt chaining to enforce narrative consistency across 9 distinct genres, maintaining coherent plot development through a multi-step generation process.

Live Demo

PlantyyGo | Plant Detection Web App

Led the technical architecture of the PlantyyGo platform. Oversaw website operations, managed the technical team, and conceptualized interactive features to ensure high performance and scalability.


6. Certifications & Coursework


7. Archive (Earlier Works & Projects)

EZTime (Apparent Solar Time Calculator): Built a research tool utilizing Python and geospatial technology to calculate apparent solar time based on user longitude. Corrects for Earth's elliptical orbit to enable data collection on solar time and biological clock synchronization.

Save The Earth Programme (STEP): Served as Web Developer, managing the technology landscape, deploying a dynamic promotional blog, and designing interactive digital games for environmental webinars (Oct 2021 - Dec 2021).

Middle School Public Speaking: Delivered a technical presentation on Phishing Websites and cybersecurity to 250+ peers. Nominated for Science Buskers (AY 2021-22) for scientific communication.