Laksh Goyal

CS at UC San Diego. I build backend systems, AI infrastructure, and developer tools.

I work across the product and infrastructure, profiling bottlenecks, improving latency and throughput, and testing whether systems recover when things fail.

Explore my work
2x Founding Engineer, Pylon and Inovasity 15 hackathon wins Open source contributor UC San Diego

Selected work

Leaf

macOS application

Find files by what you remember about them.

Search names, contents, file types, and dates in one native Mac interface. See matching lines and PDF pages before opening a result.

Source code
Swift, SwiftUI, AppKit

A shorter path to the right file

Filename, content, type, and date filters share one search. Results bring the matching context into the interface so choosing a file takes fewer steps.

Download Leaf for Mac

Native where it matters

SwiftUI and AppKit provide the Mac interface, with matching lines and PDF pages displayed alongside file results.

Read the engineering write-up

amac

Agent control plane

Know which agents and automations need you.

Sessions, permission prompts, deliveries, and follow-up work in one place, accessible from a Mac or phone.

Source code
amac dashboard showing agents and work that needs attention
GoSessions and delivery health

A complete run history

A recovered run can hide earlier failures. amac keeps deliveries, declared cadence, and the full run log visible so the history survives recovery.

Try it locally

The seeded demo creates a week of automation activity in a disposable local log.

Open local demo instructions

Strata

C++20 storage engine

Recovery with a contract you can check.

LSM-tree storage, a checksummed write-ahead log, and a crash harness that interrupts the process inside writes.

Source code
Published crash-test matrix
11,149

process kills tested

2,223,252acknowledged writes checked
0failures in the published matrix

The failure inside recovery

Repeated crash and recovery cycles exposed a torn log tail that could corrupt a later recovery. The fix durably truncates that tail before subsequent writes.

Read the published crash matrix

What the tests cover

The matrix checks process failures, which preserve the OS page cache. Power-loss durability additionally depends on the fsync policy.

Open the engine demo

Winnow

LLM input compression

Measure what shorter inputs actually cost.

Evaluate compression through useful information retained, token savings, and end-to-end latency.

Source code
Winnow playground comparing original and compressed inputs
LLM inferenceCompression playground

The information tradeoff

Token savings are evaluated alongside the useful information that survives compression.

The full request

Preprocessing adds work before inference. The evaluation measures whether the overall request benefits.

Watch the project demo

GPU Broker

Compute infrastructure

Shared GPUs. Budgets the queue enforces.

One queue across cloud and lab capacity, with per-user budgets, checkpointing, and interruption handling.

Source code
GPU Broker sample dashboard using fictional demo data
Cloud infrastructureFictional demo data

Capacity with a budget

The club shares cloud credits and lab hardware. The broker tracks who holds capacity, what it costs, and which jobs can run within budget.

Explore the public example

The sample dashboard shows utilization, queue activity, and credits using fictional personas and demo data.

Open the sample dashboard
View all projects and studies

Teams I’ve built with

Product engineering, applied ML, and early-stage teams.

DR FinancialSoftware Engineer Intern, current
Built the first doctor-credentialing product, taking a 4-hour manual process to 2 minutes of automated checks plus 10 minutes of human review.
PylonCompany raised $100K, Bow Capital incubator
Led 3 engineers from architecture through operator field tests for a counter-drone sensor mesh, including radio capture, anomaly detection, networking, and the operator dashboard.
Open the operator demo
Inovasity3D wound imaging and analysis
Worked with 12 clinicians across 3 organizations on automated segmentation and asynchronous processing, reducing wound-outlining time by 80%.
GMI CloudAI platform and product engineering
Built Mint, used by 20 teammates. A matched batch went from 20 staff-hours to 20 minutes of review by 1 person.
Qualcomm InstituteMachine Learning Researcher
Built audio validation and training-data pipelines, increasing usable recordings from 7% to 15% across multi-terabyte wildlife archives.

Hackathon highlights

SkepticFirst place in 3 tracks at Hyphenate’s Syndicate.

PylonFirst Bow Capital Defense.

Headcount3 awards at UCSD Hack for Humanity.

Open source

SwitchMerged fix for desktop login failures.

Awkward ArrayProposed Unicode serialization fix.

Community

OpenAI Student CollectiveLead at UC San Diego

GMI Cloud AmbassadorClouder program

AWS Student Builder ClubTechnical Workshops and Projects Lead

Adobe Student AmbassadorUC San Diego