Can Research Drive Product Value? Lessons From Spam Detection
summary
Research on MMS image spam produces a deployable classifier that matches GPT4 Vision accuracy on benchmarks while slashing cost and latency. The team adapts patch‑based visual transformers and CLIP style embeddings, fine‑tunes on iteratively curated open datasets labeled with Label Studio, and runs on Kubernetes. Results show comparable accuracy to GPT4 Vision, one CPU 2 GB handles 400 concurrent requests, and 2.3–4× faster than calling the GPT4o API, making it production feasible.
AI-written summary · about
chapters (10)
- 0:00Introduction and research value
- 1:40Bridging research and product gaps
- 3:06Aligning research with business goals
- 4:30Collaborating with academia and industry
- 6:37Balancing speed and research depth
- 10:48Transformers and vision models
- 13:36Fine-tuning models for spam detection
- 15:45Proving value to product teams
- 16:49Testing with real product constraints
- 18:55Benchmarking and scaling the solution
related talks
16mProduction MLOps for Spam Detection: Real-Time Multi-Modal AI
40mAI vs AI: Creating Novel Spam and Catching it with Text Generating AI
21mBeat GPT 4 and Build the Best AI For You: Mobilizing First-Party Data with Fine-Tuning
01h 03mIt's All About AI
20mGenAI synthetic data used to fine tune a domain specific model
52mBeyond the AI Hype: What's Real, What's Next
40mMachines are Learning
23mUnveiling the Journey of NLP