Sangwon Ji
Education
University of Waterloo
B.Math, Scientific Computing & Machine Learning, Joint Combinatorics & Optimization
- GPA: 3.7/4.0
Experience
Qorsa
Machine Learning Engineer (AI Engineering)
- Cut report-agent context from ~83k to ~4k tokens and latency from 77s to 12s per section, by replacing a sequential master-LLM loop with one structured plan parse, parallel section agents and on-demand knowledge-graph context
- Made tool calling reliable on open-weight models that replied in plain text instead of calling tools: Pydantic schema-enforced output and tools executed in code, saving ~25k tokens per section
- Built permission-aware citations for generated reports: every claim links to the top 5 sources its reader is allowed to open, auth-filtered and ranked over Dgraph in a single DQL round trip
SK Networks & ENCO
AI Full-Stack Apprentice
- Built an ETL and embedding pipeline that indexed financial news into Amazon OpenSearch for semantic search
- Shipped real-time apps (FastAPI, Redis, Next.js) with sub-second WebSocket and SSE streams, and CI/CD on AWS
Republic of Korea Army, 50th Infantry Division
Team Leader, Special Duty Team
- Led a 3-member escort team in a special mission unit (VIP protection, counter-terror); top pistol marksman
Projects
The Sovereign, Solo developer
- Built a self-hosted multi-agent research lab: every night it researches a topic, stores each fact with its source in a knowledge graph, and answers with citations via Graph RAG; 47 pages became 1,010 sourced facts in one night
- Made its learning measurable: after every cycle it sits an exam written from held-out sources it is never allowed to read, and the questions it misses become the next night's research (held-out score 41 → 55)
- Measured graph quality, found 68% of entities unconnected, and redesigned ingest so only source-validated facts enter the graph; made local-model extraction 9× faster (171s → 19s) at equal quality
Catchbunny, Solo developer
- Built an MCP server that lets Claude Code and Codex test their own changes: a browser swarm reports a bug only after reproducing it in a fresh browser (9/9 error-raising planted bugs, zero false positives, no LLM)
- Stopped trusting LLM testers after a local model's 3 "confirmed" bugs all proved false: a model's report now counts only if its check fails again on replay. Claude then caught the 2 silent bugs rules can't see, for $0.75
- Ran the swarm on one asyncio event loop with lock-free coordination (4.6× throughput on 8 testers) and shrank repros from 78 to 28 steps with parallel delta debugging, exported as Playwright tests that need no model
Repo Atlas, Solo developer
- Built a local-first code explorer (one command: npm create repo-atlas) that reads a GitHub repo into a knowledge graph and renders it as five layers and a call-graph mind map
- Kept the LLM out of the structure: static analysis builds the graph, the model only names and explains it. Result: 944/946 call edges verified against source, 0 fabricated citations, 16/17 answers correct
Skills
- AI & Agents
- Claude API, Ollama, Pydantic AI, LangGraph, LangChain, MCP, Graph RAG, Evals, Langfuse
- Languages & Frameworks
- Python, TypeScript/JS, SQL, C, Bash, FastAPI, React, Next.js, Node.js, Playwright
- Data & Infrastructure
- Dgraph, ArcadeDB, Neo4j, Postgres/pgvector, MySQL, Redis, Docker, Jenkins, AWS, GCP