Meme-Maker GPT

What's next for Meme-Maker GPT

I noticed that people love sharing memes, but most of them don’t have design skills or time to brainstorm clever captions. I wanted a tool that could instantly turn any topic into a ready-to-post meme idea—no Photoshop, no overthinking.What I Learned Prompt engineering: tiny changes in a GPT prompt can swing output from cringe to hilarious.Cultural context: a meme that lands in the U.S. can flop in the MENA region. I built a tiny “cultural distance” scorer

to down-rank ideas that rely on hyper-local jokes.Latency vs. creativity: letting the model think longer (higher top_p , multi-shot) improves quality, but users want < 2 s response times. I settled on a two-tier pipeline: fast mode for instant ideas, creative mode (async) if the user swipes “more”.How I Built It 1. Data: scraped 50 k top memes from Reddit + Twitter, kept only posts with > 1 k up-votes/likes.2. Fine-tuning: LoRA adapter on llama-3-8b-instruct for 3 epochs with caption + template pairs.3. StackBackend: FastAPI + Redis cache keyed by (topic, lang, safe-mode) .Frontend: React-Native WebView so the same code ships to iOS, Android, and web.Deployment: Fly.io in ams and iad regions, autoscale 1 → 20 replicas; cold-start ~ 450 ms.4. Safety: a lightweight BERT classifier (97 % F1) filters political, NSFW, or hateful content before the idea reaches the user.Challenges Bilingual puns: Arabic / English code-switching broke early tokenizers; had to train a SentencePiece model on a mixed corpus.Template alignment: the AI often picks a caption that doesn’t fit the macro image dimensions. I added a constraint solver that maximizes

and drops suggestions with fit < 0.8.Cost: GPT-4 quality is great but \$0.06 / 1 k calls adds up. I built an A/B budget throttle: once daily spend > \$20, 50 % traffic rolls back to the fine-tuned 8 B model with cached examples

Built With

• and-a/b-logsfly.io-?-edge-runtime-(ams-+-iad-regions
• android
• autoscale-1-?-20)docker-+-github-actions-?-ci/cd
• hate
• lang
• python-3.11-?-core-logic-&-ml-pipelinepytorch-2.2-?-fine-tune-llama-3-with-lorafastapi-?-low-latency-rest-apiredis-?-sub-ms-cache-keyed-by-(topic
• safe-mode)-postgresql-?-store-templates
• sub-3-min-deployshugging-face-inference-api-?-fallback-gpu-when-self-hosted-gpus-are-hotreact-native-web-?-single-codebase-for-ios
• votes
• webcloudflare-r2-?-serve-meme-templates-at-~-0-egress-costbert-base-multilingual-?-97-%-f1-safety-filter-(nsfw