My Resume

• First-authored a paper on source-grounded course generation, currently under review at the AI-Native Academia workshop (NeurIPS 2026). It introduces Grounding Fidelity, a metric under which a generated sentence only counts as supported when it is backed by a mechanically verified quote from the assigned source.
• Built the textbook-grounding pipeline for an LLM multi-agent course generator, writing each lecture slide from retrieved source evidence rather than model memory. Verified-claim rate rose from 47.9% to 60.9% across a 14-week course against a baseline handed the complete chapter text, a 13-point gain worth roughly eight standard errors, with the automated judge validated at 91% agreement against human annotators.
• Engineered the end-to-end ingestion and retrieval stack: structured PDF parsing (pymupdf4llm), GPT-4o-mini vision converting ~240 scanned equations to native LaTeX from a 740-page textbook, hybrid BM25 + dense-embedding search fused by reciprocal rank, and per-week section binding that constrains where evidence may come from.
• Ran controlled component ablations showing that evidence control matters more than prompt instructions. Removing chapter binding alone cost 24 points and pushed the system below the full-chapter baseline. Every diagnostic run was pre-registered and backed by machine-checkable receipts.
• Extended the pipeline into a storybook mode generating fully illustrated children's books with character-consistent line art; two culturally-localized books for offline Kenyan schools cleared for pilot printing by Conservation International Kenya.
• Delivered applied AI courses to external stakeholders: a source-grounded clinical-records curriculum for Mayo Clinic with per-sentence verification receipts (now expanding into an expert validation-interface collaboration), and a 6-week AI course for wildlife conservation rangers in Kenya.
• Hardened the pipeline with an 800+ test suite spanning parsing, retrieval, binding, and verification.

• Grading coursework for CSE 572 (Data Mining), a graduate course, through the Fall 2026 semester.

• Assisted Prof. Ayca Tuzmen in delivering the FSE 100 Barrett Summer Scholars engineering camp, mentoring 40+ high-school students across all hands-on sessions.
• Guided students through Python and Raspberry Pi labs, troubleshooting hardware wiring and software bugs in real time to keep teams unblocked.
• Supported students from first lines of code to a working build, reinforcing debugging habits and core engineering problem-solving throughout the camp.

• Modernized a 15-year legacy ETL platform used by 100+ enterprise clients, building scalable workflows with React, Node.js, MySQL, and PostgreSQL.
• Onboarded 10+ enterprise clients from configuration to live production, accelerating adoption by 3x through platform generalization, client demos, and deployment support.
• Reduced manual data processing effort by 40%+ by implementing reusable SQL-based ETL logic, stored procedures, validation scripts, and automated reporting workflows.
• Accelerated release cycles by 2x and reduced deployment issues by integrating CI/CD pipelines for ETL workflows.

• Delivered 15+ client applications by building production-ready frontend systems, reusable UI components, and client-facing product workflows.
• Reduced client time-to-launch by 50% by handling deployment, domain setup, Cloudflare configuration, and AWS EC2 hosting.

• Sole engineer for the admin dashboard of a WhatsApp matchmaking platform (2,400+ members), built end-to-end on React Router 7 + Cloudflare Workers to replace the team's manual Google Sheets workflow.
• Designed the data layer around a 15-minute Sheet-to-MongoDB sync via a RestHeart gateway, with Better Auth and role-based access control.
• Built the member-signup flow end-to-end: a signing proxy for Framer forms, an approval queue that turns applicants into members, Razorpay payment webhooks, and scheduled email alerts.

• Built an AI accessibility tool that writes WCAG alt text directly into PowerPoint lecture decks, describing 380 images across 505 slides in under 10 minutes so screen readers can convey diagrams to blind students.
• Engineered a FastAPI pipeline on ASU's on-premise AIR GPU gateway using Qwen3-VL and python-pptx, with a self-scored confidence gate and a two-signal guard (slide area + distinct-colour count) that routes uncertain images to human review instead of silently discarding them.
• Across 10 university decks: 228 images described automatically, 122 correctly skipped as decorative, and 30 escalated to human review, alongside 531 other accessibility issues detected across 5 WCAG checks.

• Built an AI-powered conference companion that helps students and professionals convert networking conversations into referrals, interviews, mentorship, and job opportunities.
• Implemented voice-first capture using OpenAI Whisper for transcription and Claude API for extracting contacts, opportunities, follow-ups, and outreach drafts.
• Designed connection value scoring, QR Quick Add, relationship search, and AI-generated LinkedIn/email follow-ups using React Native, Expo, TypeScript, Supabase, and Deno edge functions.

• Built a belief-aware AI agent marketplace where agents bid on tasks based on confidence, improving task routing quality using ThinkN.
• Integrated Claude, Tavily, AutoDB, and Postgres to support real-time research, reasoning workflows, and natural-language queries over live marketplace data.

GPA: 4.17/4.00

• Coursework: Data Processing at Scale, Data Mining, Foundations of Algorithms, Statistical Machine Learning, Knowledge Representation & Reasoning, Software Verification, Validation & Testing.