Rheana Mindo

Remote · Shanghai, China · C# / .NET 8 · Static Analysis · Machine Learning

• Built the code-analysis platform. Highest-volume contributor to a .NET 8 Roslyn static-analysis and refactoring platform (1,672 of 2,996 commits): a solution-wide extraction engine emitting a typed entity graph, an architecture-metrics engine (Design Structure Matrix with Tarjan SCC cycle detection, coupling, Halstead), a 60-route REST API, and the entire 37-component Blazor WebAssembly dashboard written solo, including a hand-written query language (tokenizer, parser, expression tree) for filtering millions of code entities. Raised dead-code detection accuracy from ~5% to 96.9% and rebuilt the cycle scanner from regex to Roslyn semantic analysis, taking false positives from 88% to 0% across 457 projects.
• Architecture enforcement & platform reliability. Authored 40+ production Roslyn analyzer rules (~5,700 lines) enforcing architecture at build time across four repositories, plus a 473-line analyzer checking .csproj/.sln structure at compile time; drove build warnings 1,982 to 0 and 4,408 to 78. Became the person who unbroke the build across the estate: 71 PRs repairing a broken solution across 8 repositories, including single repairs of 4,062 and 3,202 compile errors from upstream breaking changes, and 161 errors to zero across 335 files in one day.
• ML infrastructure & production incidents. Built a native C# GPU training pipeline on TorchSharp — memory-mapped streaming mini-epoch loader, two-stage fetcher/uploader threads, bounded CPU/GPU queues — removing the Python subprocess boundary, plus NanoGPT, MinGRU and MinLSTM pipelines and a Bradley-Terry pairwise ranking trainer on 1,024-dim CLIP embeddings. Shipped an LLM-driven service generating 1,800 images/day across CPPN, SD 1.4 and Flux. Root-caused 30,368 silently failed embedding tasks, a 1000× memory waste OOM-killing the orchestration server (40 MB to 40 KB per page), and an O(N²) pagination scan degrading from 0.84s to 37.68s per page; designed the GPU prefetch architecture after thermal cycling destroyed 4 GPUs (~$12,000).

Remote · Melbourne, Australia · Python · CI/CD · WordPress Elementor · Regulated Medical Advertising

• AI automation platform. Second-highest contributor of eight on the company's internal automation platform — 1,195 commits (40% of the repository), 89 of 90 PRs merged — plus sole author of a separate 131-commit QA system. Authored 9 Claude Code skills (`/qa`, `/daily-check`, `/code-review`, `/replace`, `/weekly-review`, `/popup-setup`, `/snippet-cleanup`) and 39 Python tools, compressing 5-hour manual QA passes to 3–10 minutes (~30–100× faster).
• Ad-attribution revenue recovery. Found that the Google Ads and TikTok offline-conversion uploaders selected contacts by signup date rather than payment date, silently dropping returning patients who paid in-window — ~78% of in-window payments in the June backfill. Wrote the payment-date scanner that became the shared source of truth for the window boundary, then hardened the recovery path across seven commits: deterministic order IDs, window-invariant revenue keying, UTC-consistent scoping, and test-mode and future-date guards.
• Site Health Dashboard (sole author). Designed and shipped a Flask service on Google Cloud Run end to end as its only contributor — every file, both GCS feed publishers, all tests, the reusable deploy workflow and its four callers. Serves booking and compliance health above Core Web Vitals, backed by an LRU window cache with lock-guarded background builds and self-reloading placeholders on cold ranges; ranks Lighthouse opportunities by pages affected then savings and sorts worst-mobile-first because mobile is the booking surface. Fails fast at import when `FLASK_SECRET_KEY` is unset to prevent silent open-access deploys.
• Booking integration, compliance & site quality. Fixed misrouted GoHighLevel calendars across 3 clinics and 100+ pages; hit a Monday peak near 100 bookings, previously only seen mid-to-late week. Built the tiered AHPRA/TGA compliance scanner gating a 438-page prescription-medicine advertiser, and drove PageSpeed Insights red to green across image compression, font consolidation, JS-error remediation and a cache-layer conflict that had stopped the CDN serving.

Remote · Quezon City, Philippines · Python · LLM / AI Platform

• Production AI service (sole developer). Designed and shipped a FastAPI + LangChain microservice with REST endpoints that ingests markdown from upstream Insights and Analytics bots and auto-generates positioned dashboards: 61× speedup over the prior manual workflow (36 widgets in 35 s vs. ~36 min). Authored 100 of 115 commits (87%) and 18,040 lines of Python across 94 files; containerized via Docker and deployed via GitLab CI/CD to AWS in 6 weeks. Hit pylint 10/10, 0 SonarQube issues, and 379 unit tests at 77.8% coverage; integrated Sentry SDK for production telemetry.
• LLM agent platform (core contributor). Shipped 7 production LangGraph subagents on the SourceFit onboarding AI, including a datasource agent with a dynamic OpenAPI registry (replacing a hardcoded API list), a schema-driven data-validation agent, and a KPI-retrieval agent integrated into the dashboard-selection workflow. 58 of 145 commits (40%).

Freelance · Remote · Python · OpenCV · GRU RNN · Computer Vision

• Co-developed a Baybayin Detector that recognized handwritten characters in real time using Python, OpenCV (Lucas-Kanade optical flow with KLT tracking), and a GRU RNN model, deployed with a Tkinter interface.

Freelance · Remote · MERN · Interactive Game App

• Served as client-facing point of contact for a 3-person team building a story-based interactive game app. Handled GoDaddy deployment, coordinated feature assignments across developers, and ensured client satisfaction with the final product.

Hybrid · Manila, Philippines · Python · LangChain · AI Knowledge Solutions

• Developed an AI-powered knowledge solution optimizing search and information retrieval, including a RAG chatbot built with SQLite, LangChain, Ollama, and ChromaDB integrated with Redmine data; drastically faster than the prior workflow of manually searching through Redmine posts and re-posting to surface answers.

Undergraduate Research · UPLB · Python · PyTorch · LLM Bias Research

• Quantified religious bias in LLMs (BERT, RoBERTa, DistilBERT, GPT-2, T5-small) by developing the Rapid Religious Bias Probe (RRBP) and adapting a Religion-Equalizing Loss Function (RELF) using Python and PyTorch; evaluated bias with RBS, causal/embedding metrics, StereoSet, and CrowsPairs.

On-site · UPLB, Los Baños, Laguna · Articles · Workshops · Mock Exams

• Published 21 computer science articles, boosting the COSS page to the 2nd-most-popular section (next to CS Week) and helping grow followers to 6,000+ after my term.
• Initiated peer-driven mock exams for 400+ students and workshops with alumni and industry experts on career-aligned tech skills, bridging curriculum gaps.

On-site · UPLB Institute of Computer Science · PERN · Vercel · Render · Room Reservation System

• Spearheaded the development of an ICS room reservation system with a team of 19 developers using PERN, Vercel, and Render; ensured comprehensive documentation and proper code reviews.