Tibor Thompson

Call Radar ranks the officiating calls most likely to be disputed, so fans can find them and weigh in. Its score considers the call, the game situation and fan votes, with a win-probability model estimating how much a different ruling could change a team’s chance of winning. A decision model links social posts and news coverage to the exact play. I built it and helped shape the design. In a back-test across four NFL seasons, an earlier version of the score ranked well-known disputed calls at or near the top of their games.

Posts go through automated moderation in stages, so a moderator only reviews cases that need judgment. Clear cases are settled by rules and a classifier, with thresholds I tuned on real scores to distinguish competitive banter from threats.

Every model call goes through one gateway, with a kill switch per feature and spend caps that fail closed. A call that times out is never retried, because the provider may already have billed it.

Scores and play-by-play are pushed to the app over a websocket, with fallbacks for when the socket or the data provider fails, and we get paged if scores stall.

Authorization is enforced in Postgres with row-level security, and a database test fails if a new table is added without it. Access rules are enforced in the database rather than relying on the app to protect user data.

Much of the code is written by coding agents working in parallel, so I built the CI/CD around that, with one branch and pull request per task. CI audits check rules a reviewer would otherwise need to remember, including security checks on database functions.

Every claim in an answer was matched to the exact passage in the source records, with a confidence score, so staff could spot an unsupported claim before acting on it.

With LLM-as-judge evals on the orchestrator, a prompt or tool change could be scored and compared instead of spot-checked by reading transcripts.

With a SQL action layer, plain-English requests became database operations, and metadata filters narrowed what was retrieved.

At ICMA 2025, I ran technical demos for city managers and officials from across the US.

The platform was used by live customers, with real calls and payments going through it.

Caller and job details were extracted from the conversation directly into the proposal, with no manual entry.

I also owned the AWS infrastructure, CI/CD and test suite.

Earlier, I shipped the company’s contractor-matching MVP and presented its architecture in investor diligence.