Student product, data & AI builder
Building systems that turn complex information into tools people can actually use.

Finora • MemShield Paper • LinkedIn

AI-supported financial-literacy platform for students, educators, schools, and community programs.

• Role: Co-Founder & Product Lead
• Work: Co-built the React/TypeScript + Supabase/PostgreSQL platform across personalized learning, simulations, Finny AI, progress systems, educator workflows, and school-ready identity/access.
• School systems: Built pseudonymous/no-email student accounts, classroom seat provisioning, teacher/class flows, privacy and consent controls, and privacy-aware AI memory.
• Current footprint (Sep. 2026): 3,500+ users, 2,900+ activated users, 19,400+ completed lessons, and 300+ students across 10+ school/community cohorts.
• Engagement: 64% of Lesson 1 starters continued to at least three more lessons; 31% of students in school/community cohorts continued beyond assigned work.
• Platform: usefinora.com

Operational data product that converted fragmented public permit and property records into inspectable HVAC market opportunities.

• Role: Full-Stack Data Product Developer
• Work: Built ingestion/normalization, deterministic scoring, geographic heatmaps, property dossiers, owner/occupancy enrichment, street search, and exports.
• Scale: Processed 211,000+ permits across 163,000 properties into 52,662 ranked leads, surfacing 791 high-priority prospects.

Product and AI engineering internship focused on composing applications around fragmented business data and workflows.

• Role: Product & AI Engineering Intern
• Work: Built a unified operations workspace connecting three APIs/data sources and normalizing inconsistent fields into one interface.
• AI builder: Prototyped a natural-language application builder that mapped user requests to reusable Remix components, API connections, filters, and workflow logic; reduced median prototype setup from ~45 to ~12 minutes across 20 internal test cases.
• Human-centered reliability: Found that generated dashboards could look complete despite stale, missing, or conflicting data; added source/freshness indicators, schema validation, missing-data warnings, and human-review checkpoints, improving successful task completion from 68% to 94% across eight internal testers.
• Outcome: Resulting patterns were incorporated into later internal starter templates and demos.

Research on defenses against persistent memory poisoning in stateful autonomous AI agents.

• Role: Sole Author & AI Security Researcher
• Work: Formulated the threat model; designed a three-layer defense using cryptographic provenance, trust-weighted retrieval with temporal decay, and semantic conflict detection; built the Python implementation, evaluation tooling, tests, and reproducibility materials.
• Artifact: Published 23-page technical white paper and public research repository.

Serverless defensive threat-intelligence hub that turns public OSINT feeds into normalized, searchable data.

• Role: Creator & Developer
• Work: Built Python collectors, normalization/deduplication, versioned JSON feeds, and a Next.js/TypeScript interface with filtering and export.

International startup-style AI competition for student builders.

• Role: Co-Founder, Technical & Operations Director
• Work: Built registration, submission, participant-management, administrative, and judging systems; coordinated technical operations for 100+ teams and 300+ students across 6+ countries.

• FocusGuard: Privacy-first local computer-vision utility using YOLO + OpenCV for voluntary phone-use detection during focus sessions.

• Languages: Python, Java, TypeScript, SQL
• Frameworks: React, Next.js, Spring Boot, Django
• Data & infrastructure: Supabase, PostgreSQL, MySQL, Kafka, Spark, Elasticsearch, Docker
• AI/ML: PyTorch, scikit-learn, OpenCV, YOLO, LLM/RAG systems, AI-agent security

Selected public repositories are intended as evidence of technical work. Private production systems, institutional data, and client-specific implementation details are not published. Team- and organization-level metrics are labeled as such rather than presented as individual output.