Research
What can a quantum network promise about the entangled pairs it delivers? Manuscript in preparation
October 2026
Independent Study, Information Networking Institute, CMU — with Prof. Michael Mattarock
A per-pair promise needs a trustworthy model; random tests can certify an average.
A user cannot inspect an entangled pair and still use it. I separate two promises: a guard limits a user's risk of receiving a below-threshold pair under explicit physical and calibration assumptions, while hidden random tests certify the average fidelity of the pairs kept. Those tests can also reveal model optimism and raise the guard's safety margin, although an alarm cannot protect pairs delivered before it.
- 0 / 8.8M
- bad guarded deliveries across 2D and 3D topology simulations; descriptive count
- 83 / 95
- primary hypotheses passed; 12 failures reported
- 5 + 2
- experimental datasets re-analysed and single-device IBM processor runs
The topology studies cover square grids, a grid with diagonal links, and 3D cubes up to 1,000 nodes; their zero-bad-delivery count is a simulation result, not a hardware guarantee. A separate re-analysis of Oxford's 2022 two-node experiment gave a 95% anytime lower bound of 0.917 on kept-pair average fidelity under trusted measurement settings; the device-independent bound was 0.837. Two IBM runs tested components locally on one processor. The manuscript remains in preparation.
Differentiable quantum-chemical models (DFTBML3) In progress
Quantum chemistry independent study — Advisor: Prof. David Yaron, CMU Chemistry
I am working on a differentiable density-functional tight-binding model for learning chemical Hamiltonians. The goal is to keep the physical structure of a quantum-chemical calculation while making parts of the model trainable and checking its numerical behavior against established implementations. This work is ongoing; I will share research findings after they have been reviewed.
A hybrid architecture for post-quantum EAP-TLS In progress
August 2026–April 2027
Master's Thesis, Information Networking Institute, CMU — Advisors: Prof. Pedro Bustamante & Prof. Michael Mattarock · Reader: Prof. Patrick Tague
Post-quantum cryptographic certificates are large enough to break assumptions that classical network protocols quietly depend on. This thesis looks at one specific failure mode: oversized ML-DSA (post-quantum) certificates used in EAP-TLS authentication can trigger silent Path MTU Discovery blackholes — connections that fail without any clear error, because a packet is dropped somewhere and nothing reports it.
The proposed fix combines an edge QUIC-terminating proxy at the Wireless LAN Controller with a Quantum Key Distribution–secured wired backbone, inside a hybrid WPA-Enterprise 802.1X architecture. The thesis is organized around four testable hypotheses, simulated in NS-3:
- Handshake latency — targeting a ≥40% reduction in Flow Completion Time under elevated wireless error rates.
- RADIUS server compute offload under the proxy architecture.
- IPv4/IPv6 baseline isolation, to separate protocol-specific effects from the core result.
- Sustained QKD key-delivery capacity under high-density authentication bursts.
Thesis defense and final report targeted for Spring 2027. Simulation code and datasets will be released via a CMU-hosted GitHub repository and the CMU KiltHub data repository.
Publications
Manuscript in preparation
Rajpal, T. S., & Mattarock, M. (2026). What can a quantum network promise about the entangled pairs it delivers? Manuscript in preparation.
Develops conditional per-user guarantees for entanglement delivery, an anytime certificate for kept-pair average fidelity, and a test-driven check of model optimism. Evidence includes simulations, five experimental-data re-analyses, and two local IBM processor runs.
Code, results, and audit trail (v0.9.1) ↗
Journal articles
Rajpal, T. S., & Naithani, A. (2026). NeuroCrypt: A Neuro-Symbolic AI Ecosystem for Advanced Cryptographic Data Security and Transmission. IEEE Transactions on Artificial Intelligence, 7(1), 512–521.
NeuroCrypt pairs a rule-based cryptographic engine with a deep neural network to dynamically generate symmetric keys, aiming to make encryption harder to model and attack than a fixed algorithm. Evaluated against classical ciphers (AES, DES) and neural-network-based attacks, it showed a 25% increase in measured security resilience and a 15% reduction in key-management overhead.
doi.org/10.1109/TAI.2025.3577605
Conference papers
Naithani, A., & Rajpal, T. S. (2025). Blockchain-Based Digital Health Passport: A Decentralized Framework for Secure and Interoperable Health Data Management. 2025 IEEE Zooming Innovation in Consumer Technologies Conference (ZINC).
A decentralized health-record system built to guarantee data immutability and patient ownership through a unique identifier and OTP-based consent mechanism, designed for secure interoperability across providers at national scale. It uses a hybrid on/off-chain storage model with public-key encryption and role-based access control, aiming to keep large medical records both secure and compliant without sacrificing scalability.
doi.org/10.1109/ZINC65316.2025.11100281
Naithani, A., Rajpal, T. S., & Mistry, V. (2024). TAV-Cart: Transaction Assistance and Validation Cart System for Effortless Shopping and Checkout. 2024 International Conference on Advances in Computing, Communication and Materials (ICACCM).
A system design for a smart shopping cart intended as a lower-cost alternative to RFID-based carts: dual barcode scanners and a strain-gauge weight-validation system cross-check that what's scanned matches what's physically in the cart, an LED strip gives real-time feedback, and checkout happens directly through UPI — India's national instant-payment system — removing the need for a cashier line. This is a feasibility and system-architecture study, not a deployed retail pilot.
doi.org/10.1109/ICACCM61117.2024.11059075
Naithani, A., Jain, V., Rajpal, T. S., & Mistry, V. (2024). AI Based Goalkeeper for Penalty Shot Prediction Using RNN and Pattern Matching Algorithm. 2024 12th International Symposium on Digital Forensics and Security (ISDFS).
Built a predictive model — a recurrent neural network combined with a pattern-matching module — that learns a penalty taker's shot-direction tendencies from their history and predicts where they'll shoot next. Tested against four shooting strategies (a fixed repeating pattern, always the same side, fully random, and adaptively shooting away from the keeper's last dive) at 30 shots each: fully random shooting produced the lowest save rate of any strategy, directly showing that unpredictability is a penalty taker's strongest defense against a model trying to read them. Validated against Cristiano Ronaldo's recorded penalty history from the 2014–15 through 2017–18 seasons: the AI model saved 17 of 53 penalties in situations where the real goalkeepers on record saved 13 — a modest but real improvement, and an initial prototype rather than a production system.
doi.org/10.1109/ISDFS60797.2024.10527307
·
Code repository
