Skip to content
View lgoyal6's full-sized avatar

Block or report lgoyal6

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
lgoyal6/README.md

Laksh Goyal

I write systems software from scratch, mostly C++20, Go and Rust, and I publish the benchmarks where my own work loses. CS @ UC San Diego, class of 2028.

LinkedIn · lakshgoyal.com · lakshgoyal06@gmail.com


Where I've worked

GMI Cloud · SWE Intern, AI Platform · Jun 2026 to now Built Mint, the lead-research tool the GTM team runs on. Shipped a model-routing layer that cut inference cost 47%. Benchmarked 8 routing setups on cost per solved task.

DR Financial · SWE Intern · Jul 2026 to now Built the production pipeline that decides whether a doctor can be contracted. Raised auto-clearance from 41% to 73%. Ran the adversarial miss-rate testing.

Qualcomm Institute, Engineers for Exploration · SWE, ML Systems · Oct 2025 to Jun 2026 Built the wildlife-audio training pipeline and doubled the usable data coming out of multi-terabyte archives. Wrote the boundary-aware eval metric the group still gates releases on.

AWS Student Builder Club @ UC San Diego · Technical Lead Ran workshops on EC2, EKS, Lambda and S3. Built the shared GPU queue its 20 members submit training jobs to.


Systems

Strata · LSM-tree storage engine in C++20 Beats RocksDB 50x on synchronous commits through group commit, and I publish the workloads where RocksDB wins. Killed it mid-write 11,149 times and it recovered every one of 2.2M acknowledged writes. C++20 libFuzzer ASan/UBSan/TSan · demo

Taut · Reliable-UDP transport library 6x lower worst-case latency than the kernel at 5% packet loss, 12x at 20%. The kernel wins bulk throughput by 27x, which is in the README rather than left out. C++20 epoll netem · demo

Franq · Crash-safe message queue, Go stdlib only Durable writes scale 59x to 22,979 records/s by batching every concurrent commit into one fsync. Crashed the live server 40 times mid-traffic and still delivered all 17,990 acknowledged messages exactly once. Go write-ahead log group commit

Tollgate · Multi-tenant gateway for shared LLM keys Lets a team share one paid API key with per-member limits and revocation. Rate limiting stays exact past one replica: 300 req/s holds across 3 Kubernetes replicas through Redis Lua, where per-replica limiting admitted 3x the ceiling. Go Redis Kubernetes Terraform · demo

Basis · A ledger that argues with your broker Recomputes every position from your own trade history and reports each disagreement. Found 14 cost-basis errors across 9 real brokerage accounts. Java 21 Spring Boot PostgreSQL · live

TickVault · Full-depth order book archive Records order books across 6 crypto venues in Rust at 2.7M rows/s, capturing 100% of wall-clock time on 5 of 6 feeds and publishing the sixth as 0% verified rather than rounding it up. Rust Arrow Parquet · PyPI · demo


GPU and inference

Winnow · Quantized KV cache with a fused Triton kernel One kernel unpacks, gathers from the codebook and inverse-rotates in a single GPU launch, lifting dequantization 34x from 8 GB/s to 286 GB/s for 0.47 LongBench-E points. KV memory drops 1.97x at 8-bit. CUDA Triton PyTorch vLLM

Firecracker live migration · Moving a running microVM between hosts Firecracker's maintainers closed the request for this. A ~250-line Rust patch to the VMM adds it anyway: blackout under 18 ms against a 30 ms budget, with open TCP connections surviving the host switch. Rust Go KVM

GPU Broker · One queue across cloud, spot and a lab GPU Survived 315 spot interruptions across 100 concurrent training jobs with zero work lost. Reclaimed 16% of the club's shared credit pool from machines that were held but not used. Python boto3 FastAPI · live

decode-anatomy · Where a decode step actually goes Empirical roofline for an RTX A6000: compute-bound is 0.0% at every point, and the biggest long-context kernel is a memcpy. PyTorch Nsight


Things people use

amac · Control plane for the AI coding agents on your Mac Every session and automation on one page, on your phone, over Tailscale. Work queue with fencing, proved against SIGKILL. Go tmux Tailscale

Hacklist SF · Hackathon calendar for the Bay Area Around 100 subscribers. Seven keyless sources feed one classifier, and the pipeline refuses to publish an event it cannot date, so nobody sees a dead listing. Source Next.js Cloudflare Workers iCalendar

third-degree · Get grilled on a repo you built Paste a repo, get a guided map of it, then up to 10 questions climbing from language fundamentals to what breaks if you rename a schema field. TypeScript Next.js

LooseAPI · Dev and subscription spend from your inbox Projects cloud credit balances to zero before the account shuts off. JavaScript TypeScript


Open source contributions


Recognition

6x hackathon winner

  • Vigil | 1st, Akash track, AWS Loop Engineering
  • Locus | 1st, Convex track
  • Antibody | 1st, Runloop and Reflex tracks
  • EduForge | 2nd, Replay track, AWS Self-Evolving
  • Pylon | 2nd, BowCapital Defense

Also

  • OpenAI Student Collective, 1 of 140 U.S. leads across 70 universities from 25,000 applicants
  • Published through the Lumiere Research Program, top 10% of ~200 submissions
  • 2nd place, World Robotics Olympiad India | autonomous pipeline inspection robot

Full stack

Languages C++20 · C · Go · Rust · Python · Java 21 · TypeScript · JavaScript · SQL · Bash

Systems and debugging CMake · gdb · libFuzzer · ASan/UBSan/TSan · fault injection · deterministic replay · perf profiling · netem · epoll · io_uring · write-ahead logging · group commit · SWIM

GPU and inference CUDA · Triton · PyTorch · Transformers · vLLM · CUDA graphs · KV-cache quantization · Modal · ONNX

Backend FastAPI · Spring Boot · Flask · Node · gRPC · REST · Pydantic · async job pipelines · WebSockets · Redis

Data and storage PostgreSQL · Redis · SQLite · Supabase · Cloudflare D1 · Arrow · Parquet

Infra and tooling Docker · Kubernetes · AWS EKS/EC2/Lambda/S3/IAM · Terraform · Helm · Cloudflare Workers · GitHub Actions · pytest · JUnit 5 · Playwright · k6 · Linux

Frontend React · Next.js (App Router) · Tailwind · Zustand · Framer Motion · deck.gl · Mapbox GL · React Native

ML scikit-learn · MONAI · LightGBM · OpenCV · NumPy · pandas · SciPy · torchaudio · Optuna · Ray Tune · LangGraph · LiteLLM

Freelance

I take on a limited number of contracts. Best fit:

  • Backend and distributed systems: storage, queues, gateways, and the durability and concurrency guarantees underneath them
  • LLM and agent systems: routing and cost optimization, RAG pipelines, eval harnesses
  • ML and GPU infrastructure: training pipelines, inference serving, benchmarking that reports the losing axis

Reach me at lakshgoyal06@gmail.com.

Pinned Loading

  1. pylon pylon Public

    Open-world RF drone detection: anomaly-based counter-UAS that catches novel emitters a signature library misses entirely (100% miss to 0% on held-out novel classes), plus a self-organizing UDP goss…

    Python

  2. strata strata Public

    LSM-tree KV store in C++20 with provable crash durability: 12k-SIGKILL crash matrix, zero acknowledged writes lost. Live demo in your browser.

    C++

  3. taut taut Public

    Reliable-UDP transport + SWIM membership library in C++20 - SACK ARQ, adaptive RTO with 25ms floor, reliability classes. ~6-12x lower p99 than kernel TCP at 5-20% loss. Powers tautq.

    C++

  4. tollgate tollgate Public

    Multi-tenant API gateway for shared LLM keys: distributed rate limiting that stays exact past one replica (Redis Lua). Live demo in your browser.

    Go