RM

Systems and backend engineer

Rushil Madhu

I build the infrastructure other engineers depend on.

Computer Science at the University of Maryland, graduating May 2027. This summer I was the sole intern owning Linux modem data-path tooling at Qualcomm. Before that I ran the control plane for a fleet of 50+ edge devices at nuha, and I have been a founding engineer at Paradigm AI since 2024.

Where I have worked

Follow a packet from the radio to the cloud. I have worked at every stop.

Following the packet. Hover to hold it, or pick a stop.

silicon

Qualcomm

Made driver work possible without the hardware.

Built an emulation layer that stands in for the packet accelerator, so the kernel driver can be brought up and exercised on an ordinary machine instead of waiting on scarce boards.

  • kernel driver
  • emulation
  • C
  • bring-up
Qualcomm case study

Selected work

Four systems, and what changed because of them.

Each one starts with the result. Underneath it: the problem, what I owned, and the decision that mattered. Three have longer case studies.

Qualcomm: Modem Data Infrastructure

Summer 2026, Internship · Wireless R&D

Roughly 70 end-to-end data-path cases automated, and driver work no longer blocked on board availability.

The problem

Work on the Linux modem data path was gated on scarce physical hardware, and a large share of end-to-end checking was still done by hand. Carrier customers also needed a way to configure tunnelling without programming the accelerator directly.

What I owned

The team's only intern, owning the developer tooling for this path across the RMNET, IPA and QMI kernel subsystems, and learning a large codebase with no hand-off.

The call I made

Put a stable C SDK and CLI in front of the hardware; emulate the hardware so the driver could run without it; and drive validation from a declarative topology file rather than manual steps.

  • C
  • Python
  • Linux kernel
  • RMNET
  • IPA
  • QMI
Read the Qualcomm case study

The same driver code, two backings

on hardware

driver code (unmodified)

boundary

physical board

scarce, shared, slow to get

in emulation

driver code (unmodified)

boundary

stand-in

any laptop, seconds per run

Swapping only what sits below the boundary is what removes the hardware dependency. The code above it is identical in both columns.

  • ~70

    end-to-end cases automated

  • no silicon

    needed to exercise the driver

  • 1 file

    declares every test topology

nuha: Real-Time Pipeline and Edge Fleet

Oct 2025 – May 2026, Software Engineer

Real-time path · 1.8s budget

audio + sensors
teddy-lab
Rust + Python
context
teddy-core
Rust
spoken reply

Fleet beneath it

50+ devices

Raspberry Pi · radar-agent · RAUC

wg0 overlay
nftables

Registrar · JWT + RBAC

  • /enroll
  • /config
  • /telemetry
  • /commands

device shadow · idempotent command queue · retry / backoff · audit log

Operators: Zero Trust SSO · SSH ProxyJump · Kitting UI · hawkBit

Rollback on every OTA update

Every control call crosses the overlay, is authenticated, and is logged. Nothing is reachable from the public internet.

  • 1.8s

    speech-to-speech budget

  • 50+

    Linux edge devices

  • 80%

    less setup time

Speech-to-speech held inside a 1.8-second budget, degrading gracefully instead of stalling. Setup time cut 80%, and production updates recoverable by rollback rather than a site visit.

The problem

A spoken reply has to arrive fast enough to still feel like conversation, while the sensor and audio processing behind it is heavy and ML-bound. Separately, 50+ radar devices in the field needed secure connectivity and updates that could never leave one unbootable.

What I owned

Built both Rust services on the real-time path, teddy-core and teddy-lab, and owned the device and cloud control plane end to end.

The call I made

Treat latency as a hard budget rather than a target: every stage gets a share, and anything that overruns is dropped for that turn instead of stalling the reply. Put the ML workers inside the streaming binary through PyO3 so the handoff between them costs nothing. On the fleet side, a WireGuard overlay with least-privilege RBAC, identity issued at enrollment, and rollback on every update.

  • Rust
  • Python
  • PyO3
  • Redis
  • WireGuard
  • Supabase
  • hawkBit
  • RAUC
Read the nuha case study

Paradigm AI: Document and Cloud Systems

May 2024 – present, Founding Engineer

Legal RAG

scrape
lemmatize
embed
index
retrieve
10,000+ U.S. court cases

Diagrams

SLD / DWG upload
REST API
processed diagram
+70% revenue

Deploy

GitHub Actions
Docker
AWS EC2 / EB
Nginx
Two product pipelines and the path both take to production.

The diagram and DWG APIs contributed to a 70% revenue increase. 10,000+ cases indexed for retrieval. CI/CD in place so every service deploys reproducibly.

The problem

Electrical engineers needed single-line diagrams and DWG files processed through an API, and a legal product needed retrieval over constitutions, statutory codes and 10,000+ U.S. court cases.

What I owned

Founding engineer owning the APIs, the data pipeline and the deployment path; leads a team of 3 interns on an OCR system for Government of India documents.

The call I made

REST APIs for the diagram and DWG workflows; a scraped, lemmatized and embedded corpus feeding a RAG pipeline; containerized services shipped through GitHub Actions to auto-scaling AWS behind Nginx.

  • TypeScript
  • Node
  • Python
  • Docker
  • GitHub Actions
  • AWS
  • Nginx

  • 70%

    revenue increase

  • 10,000+

    court cases indexed

  • 3

    interns led

Tether: Reliable Binary Transport

2026, Personal project · C

Continued operating at 15% packet loss and over a 9,600-baud serial link, the two conditions under which JSON-RPC/TCP failed.

The problem

Agent tool calls carried as JSON-RPC over TCP stop making progress on constrained links: lossy radio and slow serial.

What I owned

Designed the protocol, implemented it in C, and ran the benchmark against JSON-RPC/TCP end to end.

The call I made

A QMI-inspired binary framing with a CRC-8 integrity check on every frame, explicit acknowledgements and retry logic at the transport layer, rather than trusting the stream underneath.

  • C
  • Binary protocol design
  • CRC-8
  • Reliability testing
Read the Tether case study

Frame

headerQMI-style · type · len
payloadtool call
CRC-8
Benchmark outcome
transport15% packet loss9,600-baud serial
JSON-RPC / TCPfailedfailed
Tetheroperatingoperating
Only the two conditions measured are shown. Vermilion marks the transport that stopped making progress.

  • 15%

    packet loss, still operating

  • 9,600 baud

    serial, still operating

  • CRC-8

    per-frame integrity

2nd place, VTHacks 12 · 2024

ProtectEd

Python + Swift safety system combining YOLO object detection on CCTV frames, on-device gunshot audio classification, Firebase real-time alerts and Dijkstra routing to the nearest safe exit.

Winner, HackPrinceton · 2023

Relearn

Short-form learning feed built with Swift, FastAPI, Redis and GPT-4: backend APIs for lesson generation, Redis-backed caching and an infinite-scroll mobile feed.

Experience

Where the time went.

  1. Qualcomm

    Software Engineer Intern, Wireless R&D · Modem Data Services

    Jun 2026 – Aug 2026 · San Diego, CA

    The team's only intern, trusted to own developer tooling for the Linux modem data path across the RMNET, IPA and QMI kernel subsystems.

    • Designed and built a C SDK and command-line tool that give carrier customers a stable interface for tunnel configuration, instead of programming the accelerator directly.
    • Built a hardware emulation layer so the kernel driver could be brought up and exercised on an ordinary machine, removing a hard dependency on scarce physical boards.
    • Built a Python test orchestrator driven by a declarative topology file, automating roughly 70 end-to-end data-path test cases that had previously been run by hand.

  2. nuha

    Software Engineer

    Oct 2025 – May 2026 · New York, NY

    Built both Rust services behind the product's real-time voice pipeline, and owned the device and cloud control plane for a fleet of 50+ Linux edge devices.

    • Built teddy-core in Rust: the service that turns a person finishing a sentence into a spoken reply, orchestrating transcription, the language model and speech synthesis over WebSockets inside a 1.8-second end-to-end budget.
    • Built teddy-lab, a Rust binary with an embedded Python interpreter through PyO3, so the streaming and time-alignment layer shares memory directly with the ML inference workers instead of serialising between two processes.
    • Made the pipeline deadline-first rather than best-effort: context assembly returns whatever it has when its budget expires instead of stalling the conversation, and anything late is picked up on the next turn.

  3. Paradigm AI

    Founding Engineer

    May 2024 – Present · Remote

    Founding engineer on an OCR and legal-retrieval platform; leads a team of 3 interns.

    • Designed REST APIs for processing electrical single-line diagrams and DWG files, contributing to a 70% revenue increase.
    • Scraped, lemmatized and embedded constitutions, statutory codes and 10,000+ U.S. court cases for a legal RAG pipeline.
    • Leads 3 interns building an OCR system for Government of India documents, scoping an open-ended product problem into engineering work.

  4. finish’d

    Software Engineer

    Jan 2025 – Sep 2025 · Remote

    Owned and shipped 3 production healthcare applications used by 1,000+ clients.

    • Managed releases through App Store Connect for 3 healthcare apps serving 1,000+ clients.
    • Built a ride-share service on Mapbox with real-time tracking; active riders grew 40% in 3 months.
    • Led 2 interns and delivered a Redis-backed User Hub request system for centralized user data and asynchronous backend processing, from design to production in 3 months.

Earlier

  • MascotYou AI

    Sep – Dec 2024

    Real-time React dashboards for college mental-health programs; Redis caching that cut load times 60%; iOS features shipped through TestFlight.

  • 1M1B, with Microsoft

    Jul 2022 – Dec 2023

    One of 20 selected city-wide for the Microsoft Industrial Exposure Program; led AutoAid, a crash-detection app, on a Spring Boot + PostgreSQL backend.

  • Max Planck Institute for Radio Astronomy

    Mar – Jun 2022

    Led 5 interns building analytical pipelines for astronomical light-curve data with NumPy, Pandas and SciPy.

About

What I keep coming back to.

I like the parts of a system where a mistake is expensive: a kernel driver that has to come up on hardware nobody can plug in yet, a fleet of devices on networks you do not control, a transport that has to keep working when the link underneath is dropping frames. Most of my work sits somewhere on that path from silicon to cloud.

I got here from the top of the stack. I started shipping iOS apps and hackathon projects, then a healthcare team put me in charge of releases for apps that 1,000 people depended on. At nuha I moved down to Raspberry Pis, WireGuard and rollback-safe updates, and at Qualcomm I spent a summer inside the Linux modem data path, learning RMNET, IPA and QMI well enough to build the tooling my team developed against.

What carries across all of it: I want the failure modes written down before the code is, I would rather emulate a dependency than wait for one, and I treat output from AI tools the way I treat my own drafts, as something to be verified before it ships. I am finishing my computer science degree at the University of Maryland in May 2027 and looking for teams that build infrastructure other engineers depend on.

Working set

Languages
Python · C · C++ · Java · TypeScript / JavaScript · Swift / Objective-C · Rust · SQL
Systems / kernel
Linux kernel internals · RMNET / IPA driver · QMI stack · ioctl interfaces · DMA · WireGuard · nftables
Backend / data
FastAPI · Spring Boot · Node · PostgreSQL · Supabase · Redis · gRPC · REST · WebSockets / SSE
Infra / ML / mobile
Docker · Kubernetes · GitHub Actions · AWS · Nginx · Core ML · ONNX · TensorRT · SwiftUI · React

Contact

Reach me directly.