Yash Sawant
1.5K+ followersCan't let an AI mog me
I work on AI infrastructure and the unglamorous parts of inference: latency, scheduling, evals, and agents that fail the moment they reach for a tool. From serving systems to products in production. That gap is the whole job.
When I am not doing that, I run, climb, and race on Gran Turismo.
If that resonates, I'd love to hear more about what you're building.
Work Life
Side Quests
Approach
- 01
I start with the bottleneck.
Before I draw boxes, I find what is actually scarce: tokens per second, KV cache, tail latency, or a tool call that wrecks the schedule. Success is a number we can measure under load, not a diagram that looks complete.
- 02
I design the system first.
Serving, storage, and failure domains come before the model card. Batching, queues, retries, and backpressure are the product. If those are wrong, no prompt will save it.
- 03
I go deep on inference.
When something is slow or wrong, I follow it through the stack: scheduler, kernel, network, prompt. I would rather understand one tail spike than wrap another API and call it infrastructure.
- 04
I make failure cheap.
Notebooks lie. I put serving paths and agents through real concurrency, long contexts, and bad inputs until the failure shows up in traces. Evals, replay, and rollback are how the system stays honest.
- 05
I write the tradeoffs down.
Why this batch size, why this cache policy, why we did not shard yet. Alignment is a document the next engineer can use at 2am, not a meeting nobody remembers.
- 06
I own the last mile.
Tools come and go. I use them when they earn their keep. The system design, the infra decisions, and the quality of what ships stay mine.