Software Engineer

Protheus bridges the gap between scientific hypothesis and lab-ready experiment design. It turns a research idea into an auditable protocol, material plan, line-item budget, and dependency-aware timeline.

The system runs as a supervised multi-agent workflow with three orchestrator-led stages: literature quality control against OpenAlex and protocols.io, schema-enforced protocol generation grounded in researcher-selected published protocols, and deterministic operational planning that maps materials to real catalog numbers with supplier options and phased timelines.

Every draft passes validation gates before reaching the researcher. Accepted corrections are stored in long-term feedback memory, so future protocol drafts can improve with use.

A CLI incident analysis tool that ingests logs and code, redacts sensitive data, and produces a structured Markdown incident report. Uses an Analyzer-Critic loop to generate root-cause hypotheses with evidence, plus metrics and audit traces to support debugging.

I built a layout aware document conversion platform that turns DOCX files into production ready LaTeX output through a FastAPI service, a core conversion engine, and a Next.js front end. The pipeline combines parsing, optional reference PDF generation, layout analysis, calibration profiles, and dynamic LaTeX generation to preserve structure and visual intent.

The system includes guarded user options, request level rate limiting, and optional LLM routing and Gemini enhancement. I also documented the design with C4 architecture views and a runtime lifecycle so contributors can reason about request flow and conversion stages quickly.

I built a local first memory system for coding workflows that stores state, durable decisions and evidence per repository. It is designed to work through both CLI and MCP so agent tools can retrieve or update repo context quickly.

The project includes retrieval ranking, checkpoints, repo scoping controls and extension support to keep memory useful across day to day development sessions.

I built an authenticated health-plan API where Redis handles operational reads and writes while RabbitMQ pushes indexing work into the background. This kept API writes responsive and still gave searchable nested documents in Elasticsearch.

I also added ETag based concurrency (If-Match / If-None-Match) so concurrent updates stay safe and clients can avoid stale writes.

This project was split across multiple repos, but together it forms one complete CI/CD pipeline for a cloud-native web app. I used Terraform and Packer for infrastructure provisioning, GitHub Actions for automated build/test/deploy, and Cloud Functions with Cloud SQL for email verification. It reduced environment setup time by about 75% and supported autoscaling behind a load balancer.

Built a voice-based Parkinson’s detection prototype using PCA + SVM on 23 acoustic features from 31 patients’ recordings. Achieved sensitivity 1.00, specificity 0.917, and accuracy 0.967.

A hackathon prototype that combines LLM driven financial agents, a Flask backend and an interactive trading desk style web UI. It supports market/news analysis and follow-up risk Q&A in the same session.

I also worked on data side experiments including options streaming with Kafka/MSK and AWS data integrations to test end to end assistant workflows.

An ambient reminder system for caregivers and individuals with Mild Cognitive Impairment (MCI). The system translates caregiver conversations into structured reminder intent and executable trigger logic, and reacts to real-time sensor events to deliver timely reminders.

Forecasted TSLA adjusted close using macro indicators, market factors and alternative assets, then evaluated strategy performance with rolling-window backtests.

After tuning, Random Forest test RMSE improved from 26.97 to 13.05 and XGBoost improved from 14.81 to 9.10. The backtest workflow compared model driven signals against buy and hold.

Built and compared multiple recommendation approaches on Amazon Reviews 2023 (All Beauty): collaborative filtering, content based filtering and neural collaborative filtering.

From the project report, the Neural CF model performed best with RMSE 0.3993, Precision@10 0.2419 and Recall@10 0.9671.

Designed and built a database-backed rental system with a GUI to streamline inventory management, bookings, and customer operations. Implemented secure data handling and workflows to reduce manual coordination for rentals and returns.

Developed a desktop application for humanitarian aid coordination with role-based access, a unified dashboard, and resource/fund management backed by MongoDB. Added analytics views to support operational decision-making and allocation.