Connor Dore builds AI systems. Parsers that hold their shape, agents that stop and ask, models that classify what nobody had time to measure by hand.

Dense records — medical and legal — keep their meaning in structure: numbered sections, defined terms, references to an exhibit or a prior result. Split that on length and you hand the model fragments that have lost the thing they refer to, and the accuracy goes with it. So the parser walks the document’s own hierarchy and chunks on those boundaries instead, keeping each piece with the context it depends on. Every extraction then comes back through a schema, so the shape of the output is guaranteed rather than hoped for — and every field can point at the page it came from.

Boundaries land on clauses, not on a token count

Chunking

Structure-aware

Output

Schema-validated, page-cited

Stack

Next.js · Claude API · TypeScript

A desktop control room for a crew of agents. Each one runs under its own identity with its own job and its own isolated git worktree, so several can work at once without touching each other’s changes, and every action is attributable to the agent that took it. Status is pushed by lifecycle hooks rather than polled, so the moment an agent stops and wants a decision it says so — and hands back a real terminal in the same session rather than making you start over. It began as my own remake of an open agent runtime, built the way I actually wanted to use one, and it became the thing I run every day across every project.

Isolated by default; it stops and asks rather than guessing

Agents

Identity per run

Isolation

Git worktree each

Stack

Tauri 2 · Rust · React

Only 267 of the 41,884 near-Earth objects in the catalogue have ever had their spectrum measured, and composition is what sets an asteroid’s entire value. Everything else was a guess with a shrug attached. The characterization layer predicts spectral type from the orbital and photometric features we do have, taking classified objects from 267 to 838 — more than triple — and the ranking becomes something you can actually search. Value is computed from size, mass and composition; accessibility from published Δv where it exists and a patched-conic transfer everywhere else.

Every point keeps the method that classified it

Classified

267 → 838

Catalogue

41,884 objects

Stack

Python · FastAPI · React