Dee (@deeflectcom) header
Dee (@deeflectcom) profile picture

ex-fintech product designer now building AI everything

Los Angeles, CA
Joined June 2025
Photo by Dee (@deeflectcom)Photo by Dee (@deeflectcom)Photo by Dee (@deeflectcom)Photo by Dee (@deeflectcom)Photo by Dee (@deeflectcom)Photo by Dee (@deeflectcom)Photo by Dee (@deeflectcom)Photo by Dee (@deeflectcom)Photo by Dee (@deeflectcom)Photo by Dee (@deeflectcom)
Dee retweeted
whoever made the robot catch the mouse cursor on @rerundotio cooked
33
116
21
3,286
302,816
I rebuilt my Twitter from zero after a ban. Here's exactly how I grew to 500 followers in 30 days with replies, real takes, and no cross-promo deals #artificialintelligence #largelanguagemodels...Show more
1
1
1
172
How I built a universal MCP server that wraps 56 APIs into just two tools using the OpenAPI Code Mode pattern - cutting token costs 50x.... - by @deeflectcom hackernoon.com/two-tools-56-… #api #mcp
1
1
377
Whoop is too expensive so I vibe designed a cheaper hobbyist version with blueprint
8
15
3
282
27,769
The Distribution Problem Nobody Talks About buff.ly/7Cp7T7x via @deeflectcom #AI #Design #GenerativeAI
1
1
2
76
Dee retweeted
I got tired of setting up GPU servers just to test a model. So I made Vasted → One command to launch any GGUF model from @huggingface → Auto-picks the cheapest GPU that fits on Vast → OpenAI-compatible endpoint, works with OpenCode, Claude Code, whatever → Telegram bot to manage from your phone → Non-interactive mode for agents and CI, just feed it a repo → Auto-kills after 30 min idle, or "vasted down" Ran Qwen 3.5 35B on an A100 today, 83 tok/s, 4 cents for 3 prompts
2
1
2
218
"Mom, how did we get so rich?" "Your dad switched from Cursor to Claude Code"
94
39
6
675
49,350
7
156
12
1,927
114,328
Dee retweeted
Here's a fun kids' song using those AI terms: Sing to "Twinkle Twinkle": Twinkle twinkle AI light, AGI thinks super bright (like humans). CoT steps slow and right (step-by-step). Agents act all on their own (autonomous). Wrapper makes it known (simplifies). Alignment keeps it nice (human values). Fine-Tuning makes it precise (specific data). No hallucination lies (false info). Model trained wise (system for tasks). More in part 2!
1
1
2
208