Ayush Suryawanshi
@globlyoptimized
- Senior Software Engineer (Applied AI) - Learning in Public - Looking for a remote role
Może Ci się spodobać
Tried using Qwen3-VL (32B, 8-bit) to detect checkboxes. Tweaked prompts, temperature, retries… still flaky. What actually worked? Good old OpenCV + EasyOCR — clean, deterministic. Takeaway: LLMs are great for reasoning. Classical CV still wins for binary visual tasks .
Read the VL-JEPA paper today and the paper truly justifies the hype : ~50% fewer trainable parameters ~2.8 × cheaper inference via selective decoding arxiv.org/pdf/2512.10942
If you’re building enterprise AI, you need to understand where your LLM breaks — not just where it works. Tried some jailbreak testing today with GPT-5.1. Interesting findings github.com/GloballyOptimi…
Starting 2026 with a deep dive into LLM quantization Just demystified model compression - turns out reducing LLMs from FP32→INT8/INT4 is more accessible than I thought. Code: github.com/GloballyOptimi… #LLMOptimization #MachineLearning
github.com
Quantizing_LLM/quantize_llm.py at main · GloballyOptimized/Quantizing_LLM
Contribute to GloballyOptimized/Quantizing_LLM development by creating an account on GitHub.
Considered myself NOT GOOD ENOUGH untill I resolved a bug for a big daddy league startup, their team wrote an internal library & kept switching the websocket connection on and off in the workflow for no reason. Feels like the peak time to start testing myself on open source.
United States Trendy
- 1. Eric Dane N/A
- 2. Baby Keem N/A
- 3. McSteamy N/A
- 4. Jamal Murray N/A
- 5. FRLG N/A
- 6. WE LOVE YOU TAEHYUNG N/A
- 7. Aliens N/A
- 8. #TheTraitorsUS N/A
- 9. FireRed N/A
- 10. WITH TAEHYUNG TILL THE END N/A
- 11. Leaf Green N/A
- 12. Cade N/A
- 13. Pistons N/A
- 14. LINGLING BA FRAGRANCE DIOR N/A
- 15. #LinglingxDiorFragrance N/A
- 16. Knicks N/A
- 17. Grey's Anatomy N/A
- 18. #thepitt N/A
- 19. House Money N/A
- 20. UFOs N/A
Może Ci się spodobać
Something went wrong.
Something went wrong.