Jeffrey Saelee

Austin, TX · Cupertino, CA · Remote

• Built custom LLM-powered code generation tooling and validated output quality against a baseline using similarity metrics, achieving ~30–40% improvement in generated code relevance.
• Contributed to migrating a high-volume audit application (1M+ daily transactions) from Cassandra to MongoDB by rewriting queries for API changes, building new backend services, and validating data integrity between systems, reducing query latency from 10,000ms+ to under 1,000ms.
• Built custom PyCharm and IntelliJ IDE extensions in Python and Java to automate repetitive developer setup and reduce boilerplate overhead across the team.
• Configured Grafana and Splunk dashboards to monitor query performance and track bugs across backend services, supporting load testing and production observability.

Python, Google ADK, MCP, RAG, Local LLM (Gemma), React, TypeScript

• Built a multi-agent LLM app that helps family caregivers of elderly relatives manage scheduling and medication, running a local LLM so private health data never leaves the machine.
• Engineered MCP tool routing with human approval gates and redaction for sensitive actions, where the model plans and deterministic Python validates and executes, backed by a RAG pipeline.

Python, Monte Carlo search, Imitation learning, Self-play RL

• Built a CPU-only game-playing agent that competes on the live skill-rating ladder, layering determinized Monte Carlo search and policy-guided rollouts over a heuristic evaluation core, with a never-crash safety net and p99 move time under 0.2 ms.
• Ran 45+ gated experiments (paired ladder A/Bs, SPRT, Wilson bounds) and trained imitation-learned policy priors on a 640k-row dataset mined from replay episodes, hitting a peak provisional ladder rating of 972.

React, FastAPI, Discord.py, LM Studio, CoinGecko/Kraken APIs

• Designed a crypto trading desk run by 7 autonomous AI agents that meet, debate, and vote on trades in a Discord channel, featuring an LLM intent router for live CEO directives.
• Engineered semantic memory recalling past meetings via local LLM embeddings, an autonomous risk engine enforcing stop-loss, drawdown breakers, and position caps, and a React terminal mirroring the trading floor.

Java 17, Spring Boot, React, Vite, Apache PDFBox

• Created a full-stack flashcard app featuring AI-powered note parsing (PDF/TXT/MD to cards) and an interactive 3D flip card UI.
• Implemented multi-LLM support (OpenAI, Gemini, Ollama) and cognitive science formatting to generate atomic cards following the Minimum Information Principle, secured via HMAC-SHA256 stateless auth.

TypeScript, React, Express, Gemini

• Built an AI-powered job search cockpit to directly crawl ATS platforms (Greenhouse, Lever, Ashby, Workday) and automatically discover Workday tenants.
• Evaluates job listings against applicant resumes using Gemini, scoring matches and tracking the submission pipeline.

Python, TF-IDF, Streamlit, MCP

• Built an offline "intent isolation" tool that uses TF-IDF to extract the most topically dense passages from large Markdown corpora into XML-tagged context packets, cutting prompt noise for downstream LLMs with zero embeddings or cloud calls (100% local).

Python, Subprocess isolation

• Implemented drop-in skill folders dynamically loaded and executed by LLM agents, using progressive context loading (metadata → instructions → resources) to save tokens.
• Designed sandboxed subprocess execution based on published arXiv agent research.

Python, Multi-agent AI, MCP, D&D 5e SRD

• Developed a text-based narrative RPG powered by a multi-agent AI Dungeon Master with 4 autonomous subagents (World Keeper, Faction Weaver, Encounter Architect, Lore Keeper).
• Designed dynamic story spines, anti-lock loop detection, autonomous faction projects, and integrated the D&D 5e SRD via Model Context Protocol (MCP).