2-person team, backend and system design owner (July 2026)
Built and deployed a workspace-scoped voting platform supporting single-choice, multiple-choice, and ranked-choice voting, roughly 11,700 lines across the codebase.
Authored the system design document and relational data model from scratch, then implemented 17 migrations totaling about 2,000 lines of SQL
Designed the authorization layer as 40 Row-Level Security policies across 12 tables, backed by 30 security-definer functions with search_path pinned on every one
Implemented Instant-Runoff Voting tabulation in full, including exhausted ballots dropped from both the round count and the majority denominator, and ties for lowest where all tied candidates are eliminated in the same round behind an all-tied guard
Enforced vote integrity with three structural database constraints: unique(session_id, user_id), unique(vote_id, option_id), and a unique index on (vote_id, rank)
Found and closed a real ballot-secrecy vulnerability where vote_selections could be joined back to votes through vote_id, replacing the linkage with a dense_rank() anonymized ballot_ref in migration 014
Shipped the departments feature that the original design document had scoped as phase two
Directed an AI coding agent as an engineering partner: scoped tasks, reviewed every diff, and retained architecture and security decisions
Current work: adding test coverage and CI, which the project does not yet have
AuditX, Financial Auditing Platform for Companies
24-hour hackathon, product and backend lead (September 2026)
Built a full-stack web app that audits employee expenses, receipts, and timesheets at 20 to 100 person companies, groups suspicious patterns into investigation cases, and leaves every final decision to a human admin.
Designed two services sharing one PostgreSQL database: a Next.js 15 app owning auth, pages, uploads, and notifications, and a Python FastAPI analysis service owning extraction, detection, scoring, and the investigator
Built three fraud checks: exact and near-duplicate receipts through SHA-256 and perceptual image hashing, unusual spending, and impossible or overlapping timesheet hours
Scored spending risk against each employee's and department's own history using median-based baselines, so a flag reflects that company's normal rather than a fixed threshold
Added a cross-signal rule that flags hours logged at one location while a receipt places the same employee in another city that day
Grouped related findings into investigation cases, with an NVIDIA NIM model writing a brief for each: what was found, the supporting evidence, review steps, questions to ask, and plausible innocent explanations
Kept confidence scores rule-based rather than model-generated, and removed personal data before any AI call
Enforced roles with Auth.js, re-checked server side on every protected route, scoped employee queries to their own records at the database layer, and wrote every admin action to an audit log
Pushed live notifications for holds and case decisions over Server-Sent Events
Dental Time Machine, AI Dental Cost Planner
codeLinc 11 20-hour hackathon, five-person team lead (October 2026)
Led a five-person team to ship a six-step web app that turns confusing dental treatment plans into plain answers on what insurance likely pays, what the patient likely pays, and how changing when care happens across the plan year can lower the final bill.
Architected the Python FastAPI backend around one rule, the AI talks and the engine counts: every dollar figure shown or spoken comes from a deterministic cost engine backed by 1,000+ automated tests
The engine applies the deductible, coverage by category, and the remaining annual maximum per plan year, and a timing optimizer tries every allowed schedule to find the cheapest, cutting the demo patient's bill from $2,500 to $1,975
Built the conversational assistant on Claude through Amazon Bedrock, taking intake by voice or text in English, Spanish, French, and Portuguese, proposing procedures and plan details for the user to confirm, and writing a recap of the results at the end
Wrote the shared dollar guard that checks every money amount in AI output, across each language's number formats, against the engine's results, regenerating or falling back to fixed text if the AI states a figure the engine did not produce
Detected the language each message is actually written in, so the assistant answers in the language the user typed or spoke even when it differs from their setting
Added natural voice playback with ElevenLabs, falling back to Amazon Polly and then to text, with an 8-second timeout per call and clips cached in memory only
Extended the dentist search so the assistant asks for the patient's ZIP code and shows the closest dentists farther out when none are within the chosen distance
Kept medical safety rules throughout: the app never diagnoses or decides whether care can wait, and symptom questions are redirected to a dentist
Kept the app private by design, with no login, no database, and session-only state, and set up the deployment on Vercel
Architected a modular pipeline of retriever, classifier, responder, and orchestrator components
Implemented TF-IDF retrieval over a local support corpus for retrieval-augmented generation, with inference through Amazon Bedrock at temperature 0.3
Designed an escalation decision matrix with 30+ keyword triggers routing security, financial, legal, data privacy and compliance, and account access cases to specialist human review
Built anti-hallucination guarantees: grounding checks that refuse to answer when retrieval returns nothing, plus constrained system prompts
Emitted structured output with status, product_area, response, justification, and request_type fields
Placed in the top 10% globally
About
How I work
Who I Am
Undergraduate at Livingstone College, Presidential Scholar with a 4.0 GPA. I use my hardwork, determination and resilience to analyze, design, lead and build projects. I solve problems and have fun while doing it. From front-end to back-end to interpreting data, I do it all and do not limit myself.
What I Build
From full-stack applications to AI agents, I design security into the schema rather than adding it later, and I build for the cases where correctness matters: who can access what, whether the data holds up, and what happens when the system is asked something it can't answer. I also work alongside AI coding agents, scoping the tasks, reviewing the output, and owning the design decisions myself.
Education
Education
Livingstone College
January 2026 to December 2029
Undergraduate, Computer Information Systems · Salisbury, NC
