Writing

Blog

Current writing lives on TokenCost; earlier machine learning articles live on Medium. Everything below links out to the original.

Build log

Fine-Tuning Qwen3-4B on Agent Traces: the Thinking-Budget Failure Benchmarks Don't Catch

A QLoRA fine-tune of Qwen3-4B on real agent traces: completion masking, replay mix, HumanEval+ results against the base model, and a 21% silent-failure mode in hybrid-thinking models that standard benchmarks miss.

2026-07-22

On TokenCost

Inkling by Thinking Machines: Pricing & Real Cost

What Inkling actually costs to run, beyond the headline price.

LLM pricing, token math, and cost engineering

Model cost comparisons and provider deep dives

Earlier writing

5 ML Techniques to Boost Your Model Accuracy Without Adding More Data

Discover five powerful techniques to improve machine learning model performance without collecting additional data.

Read on Medium →

Building Your Personal Portfolio Chatbot with LLMs

Unlock the power of Large Language Models to create an intelligent chatbot that showcases your professional profile.

Read on Medium →

Emotion Classification on Text using Fine-Tuned DeBERTa

Learn how to perform Emotion Classification on Text using a fine-tuned DeBERTa Model.

Read on Medium →

Sentiment Analysis on Reviews Using Fine-Tuned RoBERTa

Learn how to Perform Sentiment Analysis on Reviews Using a Fine-Tuned RoBERTa Model.

Read on Medium →