Understand the system
Find the variables, constraints, delays, and feedback loops that drive the outcome.
Enterprise AI · Applied research · Business systems
Build for the problem.
I build AI systems for pricing, credit, fraud, and operations, where a wrong decision has a real financial cost. I start by finding what actually drives the outcome, then design the system around it.
Business physics◆Purpose-built AI◆Decision systems◆Production scale◆◆◆◆◆
How I work
“Before I choose a model, I want to know what actually moves the outcome.”
S / 01
A business already has forces, constraints, and feedback loops. The job is to make them visible, then build around what we learn.
Selected research
Work, with the rough edges left in.
Experiments, implementation notes, and results that changed my mind.
01 / Model evaluation
Jev Beat the 4B Model. Then We Trained the 4B Model.
Jev beat the untrained small models. Fine-tuning changed the result, including what transferred to new safety datasets.
02 / Learning systemsI Sent a Small AI Model to School. Here’s Its Report Card.
What happened when a 4B student learned from a 14B teacher across 26 puzzles, then took the test without help.
03 / Inference systemsA Mixture-of-Experts Model Is Supposed to Use a Fraction of Itself per Token.
CUDA instrumentation showed that the usual sparse-inference story did not match what the hardware was doing.
Focus
Start with the business system.
The model is one component. The real work is understanding the operating environment and building something that holds up inside it.
Build for the job
Choose the data, model, and architecture after the problem is clear.
Make it work at scale
Deploy it, measure it, and keep it reliable as the business changes.
About
Business problems first.
AI where it helps.
Saurabh Sarkar, Ph.D., has spent more than 20 years building models and decision systems across finance, supply chain, and industrial operations. He founded Phenx to turn that work into scalable systems with clear business results.
Read the longer version →