Hindsight
Analyses IBM Bob sessions, finds wasted context and generates a corrected configuration for an A/B comparison.
A/B experiment, comparison UI, React foundation, three detectors and static CI deployment.
TypeScript · React · Vite · Vitest · GitHub Actions
- −25.9% fixed overhead
- −19.7% API cost
- 56 commits by me
Context
Hindsight was built for the IBM TechXchange 2026 hackathon by a five-person team. It reads an IBM Bob session export and identifies where the agent configuration spends context without producing value: missing project rules, tools that were loaded but never called, and paid Skills with no declared use.
The problem
Agent configurations accumulate instructions and tools, but the cost of that context is usually invisible. We set out to turn a session export into an understandable diagnosis and compare the original configuration with a corrected version.
My contribution
I built the React and Vite foundation, three waste detectors, the comparison screen and the static GitHub Actions deployment. I also ran the A/B experiment behind the product: same task, same commit and same prompt, changing only the agent configuration.
My work on the project accounts for 56 commits and 20 pull requests.
Result
The experiment recorded 25.9% less fixed context overhead and 19.7% lower API cost. The original hypothesis did not predict every effect: time and turn count regressed, and the report states that explicitly.
The experiment’s value lies both in the measured savings and in the reproducible comparison between configurations.