Computer Engineering at Purdue University · AI/ML minor · Expected May 2028
I like engineering work where correctness, observability, and performance all matter: local-first products, guardrailed AI systems, concurrent crawlers, consensus protocols, retrieval infrastructure, and privacy-conscious software. My repositories emphasize runnable examples, explicit trade-offs, automated tests, and results that another engineer can reproduce.
| Project | What it explores | Engineering evidence |
|---|---|---|
| Chronicle LSM Store | C++20 embedded LSM engine with atomic WAL batches, skip lists, Bloom-filtered SSTables, positional I/O, k-way range scans, recovery, and compaction | 11 tests plus ASan/UBSan; 294,173 reads/s and durable batches at 165,175 writes/s (three-run medians) |
| Converge Local-First Board | TypeScript/React collaborative board with CRDT conflict resolution, PostgreSQL cursor replay, bounded WebSockets, Prometheus telemetry, and offline queues | 31 tests including exact-clock tie regression coverage and 200-run convergence fuzzing; 100,000-operation benchmark at 2.229M operations/second median |
| LLM Incident Response Copilot | Guardrailed RAG service that retrieves runbooks, gathers evidence through allowlisted read-only tools, validates cited plans, and flags mutating actions for approval | 24-case offline demo evaluation: 1.00 top-1 source hit rate; 53 tests and 96% coverage |
| Concurrent Web Crawler & Search Engine | Java/Spring crawler with bounded asynchronous jobs, concurrent BFS, crawl-safety controls, BM25 search, PostgreSQL, and Prometheus metrics | 25,000-document synthetic benchmark: 17.239 ms p95 and 1.000 query Hit@10; 20 deterministic tests |
| Mini Raft Store | Dependency-free distributed key-value store with leader election, majority commits, durable logs, failover, and conflict repair | Multi-process tests kill and restart leaders/followers while verifying committed data and catch-up |
| Semantic Search & Response Platform | Local-first retrieval service using Flask, SQLite, FAISS, hybrid reranking, and source-backed extractive responses | 100,000-document synthetic benchmark: 16.940 ms p95, 12,990 indexed docs/s, and 96% measured coverage |
| Causal Trace Analyzer | Vector-clock analysis that reconstructs a causal DAG, finds concurrent race suspects, and calculates a latency-critical path | 33 tests at 94.83% branch coverage; 400-event benchmark improved from about 725 ms to 130.339 ms median |
| Luma Journal | Privacy-first SwiftUI journal using Apple Foundation Models, on-device dictation, SwiftData, and review-before-write actions | No accounts, analytics, or network layer; protocol-backed services and deterministic mapping/date tests |
| ScopeLedger | Full-stack AI workflow for evidence-backed scope decisions, human-approved estimates, and verified change-order exports | 45 tests across provenance, owner isolation, optimistic concurrency, audit replay, and provider failures; documented live-provider smoke test |
| HarFlow Studio | Browser-only HAR analyzer with waterfalls, budgets, and before/after comparisons | 15 tests; 10,000-entry synthetic analysis at 14.378 ms p95 on a local arm64/Node 24 run |
| Software Engineering Portfolio | Responsive React/Vite portfolio presenting experience, architecture-focused projects, and reproducible project results | Accessible navigation, responsive layouts, downloadable résumé, and GitHub Pages deployment |
Performance figures above come from the deterministic benchmark commands and recorded results in each project. They describe local synthetic workloads, not production traffic.
For role-specific browsing: Semantic Search and Luma Journal for applied AI/ML; LLM Incident Response Copilot for LLM systems; Chronicle and Concurrent Web Crawler for backend/SWE; ScopeLedger for full-stack; HarFlow for frontend; and RelayForge for automation and reliable delivery.
- ADT — Software Engineering Intern, AI & CX Systems: Built Python/pandas and pytest workflows for 600+ AI-assisted interactions, surfaced 25+ defects, and reduced weekly triage from 5 hours to 90 minutes.
- L3Harris — Software Engineering Intern, ML Systems: Processed 200,000+ NASA NOS3 telemetry records and improved labeled attack recall from 68% to 83% through feature engineering and error analysis.
- Alpha Net — Software Engineer Intern: Shipped JavaScript/AWS workflows supporting 50,000+ monthly sessions and Swift/Kotlin REST features that reduced response time by 20%.
| Technologies | |
|---|---|
| Languages | Python · Java · C++ · C · JavaScript/TypeScript · SQL · Kotlin · Swift |
| Backend & data | Spring Boot · FastAPI · Flask · Node.js · PostgreSQL · SQLite · FAISS |
| Systems & delivery | Linux/Unix · Git · Docker · AWS · Google Cloud · pytest · JUnit · GitHub Actions |
- Start with behavior, invariants, and failure modes—not a framework.
- Keep core logic behind small interfaces so it can be tested without external infrastructure.
- Treat safety limits, useful errors, and operational visibility as product features.
- Measure results with deterministic workloads and document exactly what those numbers mean.
- Leave the next engineer a clear architecture, setup path, and honest list of trade-offs.
I’m seeking Summer 2027 software engineering internships in backend, infrastructure, data, and ML-adjacent engineering.