Aegis Mission Intelligence
Aegis is my most recent AI/autonomy project: a simulation-first mission intelligence layer that turns sensor evidence into prioritized findings, analyst review workflows, mission memory, and structured mission reports.
What it does
Aegis takes a high-level mission request and moves it through mission planning, sensor evidence processing, proposal detection, semantic scoring, candidate ranking, analyst review, mission memory, and reporting. The goal is not just object detection. It is helping an operator decide which observations deserve attention while preserving uncertain evidence.
Why I built this
The project began as a drone-based search and perception system, then evolved into a broader mission intelligence platform that can support RGB imagery, infrared imagery, video, and future acoustic or sonar inputs.
Creative process
I designed Aegis around analyst burden rather than model spectacle. Each stage asks whether it helps preserve relevant evidence, reduce noisy review queues, or produce a clearer mission record.
Technical decisions
The system is benchmark-driven and simulation-first. I use Python evaluation pipelines, labeled datasets, capture precision, capture recall, and capture F1 to compare local triage strategies against semantic review strategies.
AI integration
AI is used as part of a review and ranking layer, not as a black box. Local proposal systems preserve candidates, semantic scoring helps clean noisy queues, and mission memory keeps uncertainty available for human review.
Benchmark-driven evaluation
| Modality | Strategy | Images | Precision | Recall | F1 |
|---|---|---|---|---|---|
| RGB | Local vehicle proposals | 500 | 50.0% | 100.0% | 66.7% |
| RGB | API semantic review | 100 | 73.2% | 95.3% | 82.8% |
| Infrared | Local hot-blob triage | 500 | 89.4% | 100.0% | 94.4% |
| Infrared | API semantic review | 100 | 61.1% | 100.0% | 75.9% |
Current direction
How it works
Mission Request -> Mission Planning -> Sensor Evidence -> Proposal Detection -> Semantic Scoring -> Candidate Ranking -> Analyst Review -> Mission Memory -> Mission Report.
What I learned
Review strategy should depend on sensor modality. RGB vehicle evidence benefited from semantic cleanup, while infrared evidence performed better with fast local hot-blob triage when the signal was already strong.
