mirror of
https://github.com/blackboxprogramming/alexa-amundson-resume.git
synced 2026-03-18 06:34:09 -05:00
Alexa Amundson
ec7b1445b5
RoadChain-SHA2048: c645c1292ab1555e RoadChain-Identity: alexa@sovereign RoadChain-Full: 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
52 lines
2.2 KiB
Markdown
52 lines
2.2 KiB
Markdown
# Alexa Amundson
|
||
|
||
**AI/ML Engineer**
|
||
|
||
amundsonalexa@gmail.com | [github.com/blackboxprogramming](https://github.com/blackboxprogramming)
|
||
|
||
---
|
||
|
||
## Summary
|
||
|
||
Cloud AI APIs are expensive and you don't own the data. Deployed 27 language models on-premise across edge hardware with 52 TOPS of dedicated acceleration — full inference sovereignty at a fraction of the cost.
|
||
|
||
---
|
||
|
||
## Experience
|
||
|
||
### BlackRoad OS | Founder & AI/ML Engineer | 2025–Present
|
||
|
||
**The Problem: AI Without Vendor Lock-In**
|
||
- Needed persistent, private AI inference without per-token API costs or data leaving the network
|
||
- Deployed 27 Ollama models (48.1 GB) across 3 Pi 5 nodes — installed 2x Hailo-8 NPUs (52 TOPS total) for hardware acceleration
|
||
- Fine-tuned 4 custom CECE personality models for domain-specific generation — models that don't exist anywhere else
|
||
|
||
**The Challenge: Thermals Kill Edge AI**
|
||
- Inference on $80 hardware generates heat. A runaway generation loop pushed one node to 73.8°C — approaching thermal shutdown
|
||
- Built power monitoring (cron every 5 min), CPU governor tuning, and voltage optimization — stabilized fleet at 42°C average
|
||
- Reduced GPU memory allocation from 256MB to 16MB on headless nodes, capped frequencies, applied conservative governors — no inference quality loss
|
||
|
||
**The Stack: From Model to API to User**
|
||
- Built Ollama Bridge SSE proxy for streaming model responses to web clients in real-time
|
||
- AI image generation hub with 4 backend agents (DALL-E, Flux, SDXL, FAL) — single API, best-model routing
|
||
- FTS5 knowledge index across 156,675 memory entries — models can search their own history across 230 SQLite databases
|
||
|
||
---
|
||
|
||
## Technical Skills
|
||
|
||
Ollama, Hailo-8 NPU, DALL-E, Flux, SDXL, FastAPI, Python, FTS5, Docker
|
||
|
||
---
|
||
|
||
## Metrics
|
||
|
||
| Metric | Value | Source |
|
||
|--------|-------|--------|
|
||
| AI Models | *live* | services.sh — ollama list via SSH |
|
||
| Model Size (GB) | *live* | services.sh — ollama list via SSH |
|
||
| Lines of Code | *live* | loc.sh — cloc + fleet SSH |
|
||
| Total Repos | *live* | github-all-orgs.sh — gh api repos (17 owners) |
|
||
| SQLite DBs | *live* | local.sh — find ~/.blackroad -name *.db |
|
||
| Docker Containers | *live* | services.sh — docker ps via SSH |
|