Farhan's Projects

GRPO Only Learns When the Rewards Disagree

• llm
• reinforcement-learning
• grpo

Finding the Induction Heads in GPT-2 Small

• transformers
• interpretability
• research

Every Line of a Small GPT, Written by Hand

• transformers
• pytorch

Keeping My Own Copy of WakaTime

• wakatime
• time-tracking
• timezones

Learning Chinese by Ear From Sun Tzu

• llm
• nextjs
• audio

Keeping Columbus's Lead-Pipe Loan Fund From Running Dry

• monte-carlo
• simulation
• public-policy

Erasing Artists From Stable Diffusion at Right Angles

• diffusion
• unlearning
• linear-algebra

Fraud Triage That Spares the LLM 97% of Transactions

• llm-agents
• fraud-detection
• anomaly-detection

O'Hare or Atlanta Tops Every Centrality Ranking

• network-analysis
• r
• quarto

Trial Emulation With a Critic That Guards Time Zero

• llm-agents
• causal-inference
• healthcare

Graphing What Biomedical Papers Mention Together

• nlp
• graphs
• biomedical

The Encoder Sets the Scale of Commit Drift

• git
• embeddings
• fastapi

A Pill Dispenser Built Around One Dose Record

• hardware
• iot
• mobile

Beating Buy-and-Hold by Holding Less Stock

• hackathon
• finance
• backtesting

An ASL and Emotion Overlay for Any Video Call

• computer-vision
• tensorflow
• sign-language

Casual Riders Keep Divvy Bikes Twice as Long

• r
• quarto
• eda

Timestamps and Hashes Did Most of My Photo Culling

• nextjs
• photography
• computer-vision