· 6 min
Adneural predicts how an ad will perform, neurologically and socially, before launch by combining Meta's TRIBE v2 fMRI encoder with a 200-agent MiroFish/OASIS simulation.
[read post ]· 5 min
Retrofitting a handwriting generator with a polar-coordinate tokenizer and a 6-layer cross-attention GPT decoder: character accuracy jumps from 82% to 96% and the baseline-drift problem disappears.
· 5 min
A step-by-step walkthrough of building a realistic LSTM handwriting generator in PyTorch: mixture-density outputs, Gaussian attention over characters, style priming, and honest failure modes.
· 5 min
I was 1 of 50 ICT technicians chosen from ~15,000 applicants to support COP-29 in Baku. Tracking 1,000+ laptops, daily patches, and what it takes to run conference IT at UN scale.
· 5 min
A small 2-layer LSTM beats a same-sized transformer at forecasting global temperatures from NASA GISTEMP. Here's why small recurrent models still earn a seat at the table.
· 5 min
Feedback after launching OmarAI: why short responses persisted, how I tuned the persona for depth, and a head-to-head of GPT-4o, Llama, MPT, and Falcon as the underlying model.
· 4 min
Shipping a chatbot UI inside my existing Next.js portfolio: why a Subframe template failed, why I split the backend into Python, and how I wired the FastAPI service into the frontend.
· 5 min
The fine-tuned GPT sounded like me but remembered nothing. I compared embeddings, multi-agent file search, and a simple JSON memory in a system prompt. Here's why the JSON won.
· 6 min
How I collapsed 20,000 messy Instagram DM pairs into ~2,000 high-quality English conversations by using GPT-3.5 as a data curator: the prompt, the batch size, and the fine-tuning format.
· 6 min
How I turned years of Instagram DMs into a fine-tuning dataset for a personal AI chatbot: Instagram export structure, consecutive-message merging, and why short prompts made the model stub-respond.
· 5 min
Three months at Payriff: fusing bank data with nontraditional signals (mobile-provider behavior, ride-hailing driver earnings, and other alternative data) to score thin-file applicants, plus a RAG chatbot that taught developers the Payriff API.
· 5 min
How I taught a computer to assemble jigsaw puzzles from a photo: piece segmentation, tab-vs-blank edge classification, and a greedy graph assembler that handles most puzzles I throw at it.