RonOS — Ronak Reddy, Software Engineer

About

RR

Ronak Reddy

Software Engineer · CS & Math @ UT Austin '28

I build systems that stay fast under load — from in-kernel latency tracers to distributed LLM training and a YC-backed bioinformatics pipeline. Founding-engineer experience shipping a marketplace to production, current research in distributed training, and a habit of measuring everything I ship.

CurrentlyQuantum Computing Initiative, UT Austin

FocusML training systems & low-level perf

Based inAustin, TX

StatusOpen to SWE internships

SystemsML InfraFull-stackLow-level perfApplied ML

Experience

Where I've shipped

Three roles. Research lab, founding team, marketplace startup.

Undergraduate ResearcherUT Austin · Quantum Computing Initiative AUG '24 → NOWAustin, TX

Researching memory-efficient LLM training via gradient checkpointing + activation compression — 38% lower VRAM on 70B-parameter models.
Built a distributed parameter server (C++ / Python / gRPC) benchmarking multi-GPU comms at 2.1× throughput over PyTorch DDP.
Profiled multi-GPU runs with Nsight Systems to isolate all-reduce/comms stalls — cut step time by over 70%.

−38% VRAM2.1× vs DDP−70% step timeC++gRPCNsight

Founding EngineerKinetikGigs · dev-gig marketplace JAN '25 → MAY '25Hybrid

Architected an ML-powered QA pipeline validating builds against acceptance criteria — manual QA down 70% across 15 projects.
Rebuilt the platform (TypeScript / React / Supabase / Node) — Lighthouse 54 → 91, page loads 3.2× faster at launch.
Real-time gig matching on Postgres + Redis pub/sub: 500+ concurrent users, sub-100ms p99 on every status event.

−70% manual QA54→91 LighthouseTypeScriptReactRedis

Full-Stack DeveloperPhenom · sports-media job marketplace AUG '24 → DEC '24Austin, TX

Re-engineered a media site into a two-sided job marketplace — listings, profiles, contracts for 200+ users by Demo Day.
MVP on ShareTribe + Stripe Connect with split-payment disbursements — $12K+ in test transactions at the VC showcase.
Shipped search + filtering by sport, role, and location — lifted applications per listing 90%+ while cutting search time.

200+ users$12K+ volume+90% appliesStripe ConnectShareTribe

Projects

Built & load-tested

Swipe through. Every number was measured, not vibes.

Y COMBINATOR · PYTHON / JS / VERCEL / NUMPY

GutGutGoose — Metagenomic Classifier

From-scratch metagenomic classifier (FracMinHash + greedy min-set-cover) for a YC-backed gut-health startup. Serverless with in-browser MinHash in JS BigInt — byte-identical to the Python reference.

0.04L1 ERROR ~1.0PEARSON r 40MB→KBPAYLOAD

DIST-SYS · GO / KAFKA / REDIS / WEBSOCKETS / DOCKER

Distributed Messaging System

Horizontally scalable Go backend with Kafka fan-out, at-least-once delivery + idempotent dedup, and Signal Double Ratchet E2EE (X3DH, ephemeral DH) with session state in Redis.

50KWS CONNS 99.95%DELIVERY E2EESIGNAL

LLM TOOLING · TS / NEXT.JS / OPENAI / GITHUB API

LLM-Powered Code Review

GitHub App posting inline PR comments via GPT-4 with RAG context. Diff-chunking + token-budgeting handles PRs up to 5,000 lines; prompt caching + batching cut API cost 60%.

81%BUG CATCH 5,000LINE PRs −60%API COST

ML · PYTORCH / TRL / QLORA / vLLM / FASTAPI

Socratic CS Tutor (Fine-Tuned)

Fine-tuned Qwen2.5-7B with QLoRA on 8K synthetic Socratic dialogues. DPO preference tuning cut solution-leakage 38%→7%; served on vLLM streaming at 45 tok/s.

38→7%LEAKAGE 72%WIN-RATE 45 t/sSERVED

SYSTEMS · C / libbpf (CO-RE) / RUST / LINUX

eBPF Request Latency Tracer

CO-RE eBPF programs hooking TCP + syscall tracepoints to measure per-connection latency in-kernel. Streams 300K events/sec to a Rust agent at <2% CPU; live histograms in a ratatui TUI.

300K/sEVENTS <2%CPU ±3%vs tcpdump

iOS · SWIFT / SWIFTUI / COREML / VISION

PT Range-of-Motion Tracker

iOS app measuring joint range-of-motion from live video via on-device 3D pose estimation. Validated within ±4° of goniometer truth; RTMPose → CoreML (FP16) sustaining 30 FPS.

±4°ACCURACY 30 FPSON-DEVICE 300MEASURES

SAAS · TS / NEXT.JS / POSTGRES / CLAUDE API / STRIPE

Rental Deal Underwriter

SaaS turning a listing address into an underwriting memo — cap rate, cash-on-cash, DSCR — in under 60s. Rent comps at 9% median error; magic-link auth, Stripe billing, one-click PDF memos.

<60sPER MEMO 9%RENT ERR 10BETA USERS

Stats

The receipts

Numbers I measured on real systems — tap any card's source app for context.

GPU THROUGHPUT

0×

vs PyTorch DDP baseline — custom C++/gRPC parameter server

VRAM SAVED

0%

on 70B-param training runs

LIGHTHOUSE — KINETIKGIGS REBUILD

BEFORE · 54 AFTER · 91

page loads got 3.2× faster ahead of public launch

STEP TIME CUT

−0%

isolated all-reduce/comms stalls via Nsight Systems profiling

WEBSOCKET CONNS

0K

load-tested on the Go messaging backend, 99.95% delivery

BUG DETECTION

0%

GPT-4 + RAG code reviewer vs. a senior engineer baseline

PROD CONCURRENCY

0+

users at sub-100ms p99 on KinetikGigs matching

ALSO ON THE BOARD

🧬 0.04 L1 error on YC metagenomics (r≈1.0) · ⚡ 300K events/sec eBPF tracer at <2% CPU · 🎯 ±4° on-device pose ROM · 💸 $12K+ Stripe Connect volume

Stack

What I reach for

Daily drivers first. Everything here has shipped something.

LANGUAGES

PythonC/C++GoTypeScriptRustJavaScriptSwiftCUDAJavaSQLRBash

FRAMEWORKS & ML

ReactNext.jsPyTorchNode.jsFastAPISwiftUICoreML / VisionHF TRL / QLoRAvLLMPrismaReduxTensorFlowNumPy / PandasOpenCV

INFRASTRUCTURE

KafkaRedisPostgreSQLDockerKubernetesgRPCWebSocketslibbpf / eBPFSupabaseAWSVercelLinux

FUNDAMENTALS

Distributed SystemsOperating SystemsComputer ArchitectureConcurrencyMemory ManagementSystem DesignDS & AlgorithmsMachine LearningSymbolic Programming

Education

THE UNIVERSITY OF TEXAS AT AUSTIN

B.S. Computer Science & Mathematics

Minor in Statistics & Data Science

TimelineAug 2024 → May 2028

ResearchQuantum Computing Initiative

FocusML training systems & GPU perf

AlsoYC-backed startup (GutGutGoose)

Hoops

Engineer by day, hooper always. Keep the ball off the floor.

0STREAK

BEST 0

tap the ball

R Ron's AI usually replies instantly