Projects

Designed and built a server-authoritative assessment and eligibility engine that scores exams, enforces timing, and gates shadowing access on rule-driven, per-student logic — replacing a client-side prototype with exploitable scoring and access gaps. Built a resumable assessment pipeline that draws and shuffles per-attempt question snapshots and issues immutable certification records that persist across retakes. Implemented data protections for a minor-heavy user base: JWT auth with concurrency guards, per-account and per-IP rate limiting, and audit logging aligned to the NY Child Data Protection Act and SHIELD Act.

A patient-facing medical assistant that schedules appointments, handles prescription-refill requests, and answers practice questions over chat and live voice. The voice channel runs on Vapi with Deepgram transcription and Claude as the reasoning model; the assistant carries the web-chat context into the call and books through a bookAppointment tool that hits the app's own API, with SMS opt-in. Next.js 15, TypeScript, deployed on Vercel.

A real-time computer vision pipeline using OpenCV and MediaPipe for live skeletal keypoint extraction, pose landmark detection, and multi-class pose classification from continuous video streams — a custom Random Forest model trained on a self-collected, diverse imaging dataset. Google Gemini provides personalised coaching and multimodal interaction.

A generative-AI-powered health application integrating voice AI, multimodal LLM interaction, and RAG-based personalised recommendation pipelines — applying foundation-model approaches to biological and behavioural health data in a research-to-product development workflow.

An end-to-end agentic IT helpdesk built as a scoped ServiceNow application. When an employee submits a request, a Flow Designer trigger creates a linked incident while an asynchronous business rule calls the Claude API through an outbound REST message to classify category and priority and generate a summary; the request is updated and the team is notified in Slack with full triage details — within seconds, with no human in the loop.

v1: a workflow-automation dashboard with an agentic chat feature for lead management, email automation, and reminders via MCP servers, with ElevenLabs voice. v2 rebuilt it around genuine conversational intelligence — a three-tier memory system (Redis working memory, Supabase session memory, Pinecone long-term semantic memory), an adaptive personality engine with intent and emotional-context detection, and a tool orchestrator coordinating Gmail, Twilio, Calendar, CRM, and market-analysis tools plus 87 MCP tools via Claude Flow. Parallel initialisation and lazy model loading cut boot time 60–70%.

Relational schemas in MySQL for sales and inventory tracking; SQL and ETL workflows for KPI reporting; interactive Power BI dashboards with DAX modelling; and demand-forecasting and trend-analysis pipelines for actionable business insight.

An empirical comparison of AI-agent-authored and human-authored pull requests on GitHub using the AIDev dataset (50 agentic + 50 human Python PRs). Human PR descriptions are ~7.5× longer on average (5,669 vs 755 characters) and carry significantly more dependency-management signals (Mann–Whitney U, p = 0.0002, Cliff's δ = −0.36); agentic PRs mention testing more often. Static analysis with radon and pylint on the changed files found no statistically significant difference in cyclomatic complexity or smell density at this sample size. Co-authored with Iyashi Pal.

A quantum-kernel SVM that swaps the RBF kernel for circuit fidelity in a 2ⁿ-dimensional Hilbert space, benchmarked honestly against classical RBF on Iris and two-moons with decision-boundary and kernel-matrix visualisations. The quantum kernel underperforms at 2 qubits (0.80 vs 1.00 on Iris) — a documented baseline with a clear limitations analysis.

A complete Grover implementation with dynamic multi-controlled-X oracle construction, multi-target search, and step-by-step amplitude-amplification visualisation. Framed for ML practitioners as a quadratic speedup for binary feature selection: 2ⁿ subsets searched in O(2^(n/2)) oracle calls.

A runnable simulation of the BB84 protocol: Alice encodes key bits in random Z/X bases, Bob measures, and the sifted key is checked for eavesdropping via the quantum bit-error rate. Without Eve the QBER is 0% and a secure key is produced; with an intercept-resend attacker it jumps to 24.49% (theory: 25%) and the protocol aborts. Runs on AerSimulator or real IBM Quantum hardware with one flag.

Designed and implemented a generative adversarial network (GAN) pipeline for converting natural-language descriptions into motion imagery, applying deep learning, data augmentation, and generative techniques on cognitive and neurological imaging-adjacent datasets — designed to assist individuals with aphantasia, particularly those with ADHD.

A comprehensive review of vulnerabilities and threat vectors across IoT device ecosystems — attack surfaces, authentication mechanisms, firmware security, and network-layer exposure — proposing a layered security framework for manufacturers that addresses privacy, integrity, and availability across the device lifecycle.