Kannan MURUGAPANDIAN

AI Systems Pre-University Researcher, DPS International School, Singapore

Kannan's research interests lie at the intersection of large language model evaluation, educational technology, and computational cryptography. His work examines the epistemic limitations of retrieval-augmented generation systems, the economics of distributed inference, and the design of auditable AI systems for pedagogical application. He is concurrently engaged in competitive informatics and cybersecurity instruction.

1. Education

DPS International School, Singapore Class 12, Indian School Certificate (ISC) | 2016 – 2027 (expected)

• Sole recipient, by merit qualification, of his school's placement in the competitive track of the National Cybersecurity Olympiad (NCO).
• Quarter-finalist, National STEM Championship (NSTEMC), 2025, as part of a four-member research team.
• Top 10 finalist (Singapore), SIM–LSE Data Challenge.
• Admitted to pre-collegiate academic programs at Georgetown University, Northwestern University, Rice University, and Dartmouth College.

2. Teaching

Competitive Cybersecurity 101

Instructor, Cybersecurity Club, DPS International School | 2026 – present

Designed and delivers a foundational curriculum preparing students for competitive entry into the National Cybersecurity Olympiad (NCO) and International Cybersecurity Olympiad (ICO). The course develops students from first principles through competition readiness across four domains of study:

• Python foundations and CTF scripting, with emphasis on automation, parsing, and exploitation techniques
• Cryptography, surveyed from classical substitution ciphers through post-quantum theoretical frameworks
• Digital and network forensics, encompassing memory analysis, packet capture (PCAP) interpretation, and disk imaging
• Web exploitation methodology

Instructional materials, including lecture slides and challenge sets, are released under a CC BY–NC 4.0 license for non-commercial adaptation and reuse by educators and CTF communities. The inaugural module, "Introduction to CTF and Python Foundations," is publicly available.

Course Repository

3. Publications and Invited Talks

Listed in reverse chronological order

From Semantic Reuse to Auditable Tutoring

Research Essay, 2026

This essay develops a governance-oriented extension of the Crowdsourced Semantic Cache Network (CSCN), addressing the computational and institutional barriers to scaling AI-based tutoring. A multi-stage pipeline is introduced comprising safety screening, privacy transformation, and intent canonicalization, which relocates the unit of semantic reuse from unstructured conversational text to discrete, pedagogically aligned "tutoring nodes." The resulting architecture reduces marginal inference cost through semantic caching while preserving the transparency and auditability required for institutional oversight of Socratic AI interaction.

Essay

Judgement Day: Benchmarking "Black Box" LLMs With Open Legal Datasets

Open Source Summit India, 2026

This paper identifies and names Statutory Myopia, a previously undocumented failure mode in retrieval-augmented generation (RAG) systems in which language models systematically privilege retrieved statutory text over pre-trained knowledge of superseding or conflicting legal authority. Three frontier models — GPT-5, DeepSeek R1, and Gemini 2.5 Pro — were evaluated using an original harness constructed on the LegalBench corpus (three doctrinal task subsets, 239 cases, 4,302 reasoning traces) under No-RAG and Statute-Only RAG conditions.

• Incomplete retrieval produced a 9-percentage-point decline in aggregate accuracy (0.87 to 0.78).
• The Hearsay task exhibited a 21-percentage-point decline when the superseding precedent Crawford v. Washington was withheld from the retrieval context.
• The study introduces a reproducible asynchronous evaluation harness employing SHA-256 cryptographic caching and a blacklist protocol for controlled retrieval experiments, with implications extending to any hierarchically governed knowledge domain, including clinical guidelines, API deprecation schedules, and regulatory compliance frameworks.

Slides | Recording

Cracking the Code: From Caesar Ciphers to WhatsApp Encryption

STEAM Club Lecture, 2026

This lecture traces the historical development of cryptographic method from the Caesar cipher (circa 100 BC) to contemporary symmetric and asymmetric systems. It establishes the vulnerability of substitution ciphers to frequency analysis, develops the RSA cryptosystem (Rivest, Shamir, and Adleman, 1977) through a worked derivation grounded in modular arithmetic and integer factorization, and concludes with an applied analysis of the Signal Protocol underlying WhatsApp's end-to-end encryption (Curve25519, AES-256, HMAC-SHA256), including its distinction between content confidentiality and metadata exposure.

Slides

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

Preprint, 2026 Murugapandian, Kannan

This paper formalizes a four-layer architecture (CSCN) for reducing the operational cost of large-scale LLM inference through a globally shared semantic vector database, a demand-triggered validation gate employing an LLM-as-judge mechanism, and a freemium economic model that converts user expenditure into a compounding shared knowledge resource. Closed-form cost models are derived across the empirical range of cache hit rates. Under conservative production assumptions ((N = 1{,}000{,}000) queries/day, (H = 0.40)), the architecture yields a 39.9986% reduction in raw inference expenditure; under optimistic assumptions ((H = 0.67)), reductions reach 66.9977%. Validation calls are shown to be approximately 2.8827 times less costly than full generation calls, supporting a gross margin of approximately 27.5% under paid-tier operation, with the system exhibiting positive network externalities as the shared knowledge base expands.

DOI: 10.5281/zenodo.19401231 | Non-technical summary

Reasoning Over Recall: Why Statute-Only RAG Degrades Performance on Doctrinal Legal Tasks

Preprint, 2026 Murugapandian, Kannan

This study interrogates the prevailing assumption that retrieval augmentation uniformly improves factual grounding in legal reasoning tasks. Across 239 cases drawn from three LegalBench task categories, statutory retrieval unaccompanied by case law is shown to degrade ensemble majority-vote accuracy by 9 percentage points (0.87 to 0.78), an effect most pronounced under a Textualist judicial persona (−12.1 percentage points) and within the Hearsay task (−21 percentage points), where controlling precedent was withheld. The findings indicate that incomplete retrieval can be more detrimental to legal reasoning accuracy than the absence of retrieval altogether.

DOI: 10.5281/zenodo.19522950 | Non-technical summary | Technical appendix

Democratizing Personalized Learning Using Open Source AI

FOSSASIA Summit, 2026

This talk situates Bloom's Two Sigma Problem within the contemporary AI landscape, arguing that dependence on closed-source language models introduces material privacy risk in educational contexts. A "Glass-Box Learning Architecture" is proposed, employing locally hosted 7B-parameter open-source models fine-tuned via LoRA, augmented with verified retrieval and secondary guardrail classification, to enable cost-efficient, pedagogically Socratic AI systems.

Slides | Recording

Introduction to Digital Forensics

DPS International School, March 2026

A pedagogical presentation delivered to junior students on the fundamentals of digital forensics as applied to Capture-the-Flag competition, addressing file signature analysis, PCAP network traffic interpretation via Wireshark, LSB steganography, and disk imaging using Autopsy.

Slides

Artificial Intelligence in Education

Medium, June 2025

An analytic essay examining structural inequities in contemporary education systems through a review of over eighteen peer-reviewed studies on teacher quality disparity and curriculum bias, proposing AI-mediated personalized learning as a scalable corrective.

4. Professional and Research Experience

Research Team Lead

The Junior Academy, New York Academy of Sciences | September – November 2025

• Appointed Team Lead within a globally competitive selection process (acceptance rate below 7%), coordinating a distributed research team across the United States, Singapore, and India in the development of a renewable energy proposal.
• Directed the construction of Python-based simulations integrating real-time meteorological and Gemini API data to model energy efficiency across diverse climatic conditions.
• Formulated a hybrid energy system incorporating self-cooling floating solar arrays and building-integrated photovoltaics, demonstrating a 15% efficiency improvement over conventional grid infrastructure through data-driven modeling.
• Co-authored a feasibility study on a chemical storage mechanism capable of sustaining municipal power supply for twelve hours post-sunset.

Slides

Co-Founder and Co-Head

Earth Status, Youth Climate Action Initiative | December 2021 – October 2024

• Co-founded a youth-led environmental education initiative and directed its technical development, including the launch of a mobile application on the Google Play Store and accompanying web infrastructure.
• Oversaw digital communication strategy across multiple platforms and coordinated a cross-functional content team.

Museum Ambassador

HeritageSG, National Heritage Board | August 2025 – present

• Serves as a public-facing interpreter at Singapore's national museums, facilitating exhibition tours and interactive learning engagements for diverse audiences.

5. Honors and Distinctions

6. Selected Technical Work

LeginAI — Multi-LLM Reasoning System. A nine-agent consensus architecture employing structured voting logic to mitigate hallucination in legal verdict generation, grounded by a retrieval-augmented pipeline constrained strictly to uploaded statutory text.

DonumAI — AI-Mediated Instructional Platform. An educational technology system addressing disparities in instructional quality through comprehension-adaptive lesson generation, implemented via a Flask backend integrating the Google Gemini API. Demonstration

Learn with DonumAI — Curated Knowledge Podcast. A serial audio program distilling scientific, historical, and interdisciplinary subject matter through AI-assisted curation. Podbean

MittentisAI — Narrative Generation Engine. A modular prompt-chaining system maintaining narrative coherence across nine literary genres through multi-stage generation. Demonstration

PlantyyGo — Plant Identification Platform. Directed technical architecture and platform operations, with responsibility for system performance, scalability, and feature design.

7. Certifications and Coursework

• CSARC-certified in Arduino microcontroller programming (Grade 2).
• Proficient in Python, C++, C, Java, Flask, Django, Git, and Linux systems administration.
• Coursework completed: Introduction to Raspberry Pi (Eduonix); Intermediate Python 3 (Udemy); Linux for Absolute Beginners (Eduonix); White Hat Hacking and Penetration Testing (Eduonix); Git Fundamentals.
• Logic Coders: Core Coding Logic Program, Levels 1 and 2.
• Environmental certifications: WWF Certificate of Completion; Sembcorp Marine Green Wave Environmental Care Project.

8. Earlier Works

EZTime — Apparent Solar Time Calculator. A geospatial research tool computing apparent solar time from user longitude, correcting for orbital eccentricity to support studies of biological clock synchronization.

Save The Earth Programme (STEP). Served as web developer, maintaining promotional web infrastructure and designing interactive materials for environmental webinars (October–December 2021).

Public Address on Cybersecurity. Delivered a technical address on phishing methodology to an audience exceeding 250 students; nominated for the Science Buskers program (AY 2021–2022).