SUMUKH ARADHYA · LOS ANGELES

I build systems that ship without me.

LEAD ENGINEER AT ONETERA, REPORTING TO THE CEO · PUBLISHED LLM RESEARCHER · LEADS TECHNICAL DISCUSSIONS WITH CLIENTS

THE EVIDENCE FOLLOWS

Fog over Tomales Bay at dusk, Inverness, California

min → 10 s → <1 s

FULL-TURN RESPONSE TIME · PROGRAM ORCHESTRATOR

219 / 0

SCENARIOS VERIFIED / BAD MERGES · PROVING GROUND

10+

CITIES IN PRODUCTION · TECHNICAL LEAD FOR MUNICIPAL AND STATE CLIENTS

LENOVO

PARTNERSHIP DRIVEN END TO END, PRODUCTION INFERENCE ON PARTNER HARDWARE

SELECTED WORK

01 · ONETERA · 2025

Program Orchestrator

The runtime behind every resident interaction.

A LangGraph engine: twenty-six nodes, eighty-four typed directives, Redis checkpointing. It runs every application experience across the product suite. The foundation came out of a month of brainstorming with the CEO and building alone; then the team came aboard, and together we drove full-turn responses from minutes to ten seconds, and to sub-second on deterministic actions.

MINUTES → SUB-SECOND

※ One month alone, then a team sport. The sub-second part happened out of stubbornness.

02 · ONETERA · 2025

Software Factory

An engineering platform that writes most of our software.

An autonomous platform: a Python runtime the size of a small company, 118 agent skills, 34 parallel git worktrees. It moved us from hand-written features to machine-generated commits merged through compliance gates. A three-engineer team now sustains sprints of fifty-plus shipped work items.

3 ENGINEERS · 50+ ITEMS / SPRINT

※ A massive number of commits merged. I have personally read a fraction of them. That is the point.

03 · ONETERA · 2025

Proving Ground

What makes autonomous merging safe.

Every feature must clear real-browser scenario verification and produce a sealed, content-addressed evidence bundle before merge, with verdicts that keep infrastructure faults from being scored as product failures. Two hundred nineteen scenarios, proven live on a full sprint.

219 SCENARIOS · 0 BAD MERGES

※ Its first live sprint caught five real faults. None of them were the failures I designed it for.

04 · PUBLISHED RESEARCH · 2024–25

QDyLoRA + Contrastive RAG

Making near-SOTA code review cheap to train.

Fine-tuned Llama-3 3B with QDyLoRA to near state-of-the-art code-review performance at a fraction of full fine-tuning memory cost, then designed a contrastive RAG framework for context compression that improved inference quality on open-source code-review benchmarks.

NEAR-SOTA AT 3B PARAMETERS

※ A 3B model on a grad-student GPU budget. The constraint turned out to be the feature.

THE FOUNDATION

The majority of Onetera's product surface area runs on an architecture I drew. The part I'm proudest of is that the team now builds on it faster than I ever could alone.

THE RECORD

SINCE 2025