Portfolio

A hand-built ReAct agent loop running on a local Ollama model, with no agent framework underneath: the client parses the model's replies, decides when to call a tool, and handles the failure modes (bad tool names, missing input, off-format replies) with a bounded retry instead of failing silently.

Tools now run behind a real MCP server rather than in-process — the client launches it as a subprocess over stdio, discovers its tools at startup, and calls them through an MCP client session. I'm building this to understand the ReAct loop and the Model Context Protocol from first principles, as hands-on practice for the MCP and tool-using-agent sections of my own learning plan, rather than to ship a product. Work in progress.

Research application for building weighted portfolios, retrieving dividend adjusted historical prices, and comparing results against market, gold, bond, savings, or inflation benchmarks. It supports buy and hold strategies plus monthly, quarterly, and annual rebalancing.

The engine models transaction costs, reports risk and attribution metrics, and uses a moving block bootstrap for confidence intervals. A 52 test suite checks metric calculations, invariants, benchmark compounding, data ingestion, and API behavior.

Complete point of sale platform with customer ordering, staff roles, and inventory management. Built a normalized PostgreSQL schema with triggers, transactional order processing, and audit logs. Implemented a responsive JavaFX UI with role based access, receipt printing, and admin dashboards.

Backend uses a DAO layer over JDBC, batch inserts, and parameterized queries for safety and speed. Deployed on AWS with environment configs and GitHub Actions for CI build checks.

Production style event driven analytics pipeline that ingests high volume service events, streams them through Redpanda with Kafka compatibility, computes rolling KPIs, persists results to PostgreSQL, and visualizes health in a near real time React dashboard.

Built as a clean multiservice repo with containerized local infrastructure and a streaming worker for aggregation and durable KPI writes.

Internal tool that turns raw event level data into business KPIs: revenue, activity, daily trends, and event type breakdowns. Strict role based access control gives admins platform metrics while users see only their own.

Node and Express backend with JWT auth and enforced RBAC, backed by PostgreSQL with indexed time series queries. React keeps the dashboard clean and readable.

Implemented a Windows traceroute tool that sends ICMP probes with controlled TTL values, resolves hop names asynchronously, and prints ordered path results with RTT timing and error handling. Includes raw socket mode plus a Windows ICMP Helper API fallback.

Batch mode traces many targets with worker threads and a topology aware smart mode that reuses stable prefixes. The public repo includes an academic integrity notice.

Built a custom reliable sender over UDP with connection setup, packet sequencing, ACK processing, sender window flow control, retransmission timers, RTT/RTO updates, duplicate ACK fast retransmit, and graceful FIN close.

The sender coordinates producer and background ACK threads to keep the transfer window full while tracking throughput and checksums. This originated from CSCE 612 coursework; the public repo includes an academic integrity notice.

Implemented a command line DNS client that builds raw DNS queries, sends them to a chosen resolver, and decodes answers, authority records, and additional records.

Supports A and PTR lookups, compressed name parsing, retries, timeouts, and sockets across platforms. The public repo includes an academic integrity notice.

Multi objective routing that balances time, fuel, and operational risk. The engine combines live conditions with a news derived risk score to adjust edge weights in real time. A Transformer forecaster and PPO support routing under uncertainty.

A Streamlit UI compares fastest, cheapest, and safest routes with short LLM rationales. Cached data, reproducible configs, scenario replays, and runtime monitoring support analysis.

Backend first semantic search MVP that embeds images and text in the same CLIP vector space and returns ranked results with cosine similarity. Exposes clean endpoints for ingestion, search, health checks, and metrics.

Uses a simple in memory vector store, containerized development with Docker Compose, and basic observability dashboards.

Controlled benchmark comparing a compact CNN to a Vision Transformer under limited data regimes. Tracked scaling behavior across dataset fractions, seeds, and augmentations; exported metrics and checkpoints for reproducible comparisons.

Includes training curves, calibration checks, and error analysis to show when transformers win and when they do not.

Classic VGG19 neural style transfer with a streamlined UI. Users tune content and style weights, TV loss, and iteration budget. A coarse to fine schedule and multiscale Gram features produce crisp textures.

Includes deterministic seeds and progress snapshots. Runs on CPU or GPU, with large images tiled to keep memory bounded.

Based on "A Neural Algorithm of Artistic Style" →

Transparent fake review classifier that separates CG (computer generated) from OR (organic) reviews. It stays lightweight and explainable without opaque embeddings.

Uses manually designed features such as length, suspicious phrase count, and uppercase word count, then trains an Adaline model with batch gradient descent.