Engineering Questions
Each project is indexed by the question that caused it to exist. The implementation is evidence; the repository name is secondary. Measurements appear only when they were recorded during the work.
Visual computing
| Work |
Question |
State |
| Medical video systems |
How can rare epileptic events be detected in long, constrained clinical video? |
Research completed |
| BrowserPose |
Can action recognition run privately in a browser without a backend? |
Completed |
| Fast Table Detect |
How far can classical CV go before a document model is necessary? |
Completed |
| Segmentic |
Which pruning level improves edge throughput without destroying segmentation quality? |
Completed |
| DepthEye12 |
Can monocular depth add useful distance estimates to ordinary object detection? |
Completed |
| Trident-WM |
Can a robot predict future visual states in a compact latent space? |
Completed |
Runtime and infrastructure
| Work |
Question |
State |
| EchoPulse-CPP |
What does an ESN runtime look like without Python or PyTorch at the edge? |
Completed |
| HotStart |
Can loading and execution overlap to improve LLM time-to-first-token? |
Negative result |
| More projects |
Can vLLM have a small, Ollama-like operational interface? |
In progress |
| More projects |
Can test-time latent computation extend a frozen language model? |
In progress |
Applied ML
| Work |
Question |
State |
| VoiceTrack |
Can meeting transcription remain local and still identify speakers? |
Completed |
| More projects |
How can BERT adaptation and serving fit a smaller training budget? |
Completed |
Professional systems
The professional ledger also includes a 17,000-line ACL migration, multi-GPU VLM serving, OCR infrastructure, secure clinical data collection, a revived Haskell/Jupyter system, and reproducible bare-metal deployments. These are documented in Systems Works and Research.