Arjun Aggarwal

I’m a founding engineer at Lightfield (opens in new tab) in San Francisco, where I work across the APIs, agent tools, workflows, and product surfaces behind our CRM.

This site is where I explain what I’m learning and keep track of restaurants, cities, films, music, and things I don’t want to forget.

Selected work

A few things I've worked on

all work

Lightfield

2025–present

Founding Engineer

I work across the product and systems behind Lightfield's CRM. I've helped turn product operations into a public API and Python SDK used by our own agent, built tools for creating and editing CRM tasks, and worked on workflow automation, human review, notifications, and core product surfaces. The recurring challenge is translating an ambiguous request into a change that is understandable to the user and correct in the product.

Amazon Web Services

2024

Software Development Engineer Intern

I built reporting and reconciliation infrastructure for roughly ten million financial events a month. The system used SNS, SQS, Lambda, DynamoDB, EventBridge, S3, and CloudWatch to trace events across services and preserve reporting completeness through retries and partial failures. The point was not throughput by itself. Financial data had to remain trustworthy when the underlying delivery system was asynchronous and imperfect.

Capital One

2023

Machine Learning Engineering Intern

I worked with a roughly 900-million-edge graph representing relationships in card data. I rewrote motif queries with GraphFrames and improved their performance by about six times. The project taught me that a graph is useful only when its relationships can answer an operational question quickly enough to affect the product; scale alone is not the interesting part.

Selected writing

Recent writing

all writing

What building agent actions inside a CRM changed about how I think about chat, workflow state, and human control.

From the log

Lately

full log

Jun 7 · thought

Agents need workflow state: The best AI products don't feel like chatbots. They feel like someone quietly cleaned up the workflow graph behind the scenes.

Jun 7 · album

Mk.gee, Two Star & The Dream Police: Night-driving music for a city you don't live in anymore. Still growing on me.

Jun 6 · film

Heat: Competence, loneliness, and LA at night. The diner scene still does more with two men talking than most films do with everything.

Jun 6 · restaurant

Trèsind Studio: Technically ridiculous. The pacing and the room mattered as much as the food.

Jun 5 · link

Agent evals and fake precision: Saving this because it explains why agent evals often become fake precision: clean numbers on a benchmark that doesn't resemble the real workflow.