Ayush Ranjan
Notes on machines that learn and systems that don’t.
Vol. III, Issue 14 · Bengaluru · Updated 21 May 2026 · An engineer’s broadsheet
The lede

Building ML systems
for the boring middle layer
between training and the bill.

I’m a machine-learning engineer who spends most of my time on the parts no slide deck loves: inference latency, cluster economics, and the slow craft of making a software system you can reason about. This is where I work in public — in essays, projects, and a quiet reading log.

· Start with the essays →
Currently writing
A small paper on cost-aware batching for mixture-of-experts inference. Mostly drawing.
Currently reading
Ousterhout, again. The second half hits harder this time around.
Open to
Quiet conversations about ML infra, system design, and books.

Recent essays

The archive (47 essays) →
  1. Field notes
    Apr 22 · 11 min

    Cost-aware autoscaling for inference, or: why p99 lies

    Notes from a quarter spent watching a GPU fleet melt money. A small model of demand, a smaller model of cost, and the gap between them.

  2. Architecture
    Apr 03 · 9 min

    Your microservice is a distributed monolith

    On call boundaries vs. data boundaries. The graph that matters is the one you can’t see in the org chart.

  3. Notes
    Mar 18 · 6 min

    Quantization, with a spreadsheet

    INT8 vs FP8 vs the rest, with a working sheet you can copy. The point is to feel the bits, not memorize a table.

  4. Case study
    Feb 29 · 8 min

    When pgvector is enough

    A two-million-row case study, and the thresholds I use before reaching for a dedicated vector store.

From the workshop

All projects →
  1. Subroute

    shipping

    A tiny request router for sharding LLM traffic by token-budget.

    ~600 LOC of Go. In production behind a 4-node inference cluster.

  2. Mistlake

    shipping

    Cold-storage layout for inference logs that survives a 90% read-skew.

    Cuts the S3 bill by ~38% on our shape. Rust + Parquet, single binary.

  3. pgvec-bench

    paused

    A no-frills benchmark suite for pgvector vs Qdrant.

    Mixed read/write, skewed dimensions, realistic recall curves.

From the reading log

The full log →