Sagar Patel

Systems & AI Engineer

Open to systems, robotics, edge AI, forward-deployed, and embedded software roles

$~/|

Systems & AI engineer building embedded Linux, edge AI, robotics, and high-performance networking systems that have to run fast, recover cleanly, and ship under real constraints.

4+ years shipping production systemsEmbedded Linux + CUDA + DPDK + ROS2Open to systems, robotics, and edge AI roles

View Hiring Story Work I'm Proud Of Resume

Quick links/blog /uses /projects /hiring

recruiter snapshotopen now

Best fit

Embedded Linux, edge AI, robotics, and networking systems with direct production ownership.

Location

Riverside, CA

Open to

Systems roles • Robotics / Physical AI • Edge AI / inference

Hire signal

End-to-end owner with a systems-first bias: observability, rollback paths, latency discipline, and real deployment constraints.

sagar-os — bash — 80×24

Welcome to SAGAR_OS. Type 'help' for available commands.

specializationsrecruiter scan

Embedded Linux & BSP

Edge AI inference

Robotics pipelines

Networking / DPU / DPDK

Forward-deployed systems

selected_impact

Proof points recruiters can scan in under ten seconds.

31%

signal

throughput gain on logistics edge gateways

Packet processing, DMA paths, and interrupt handling tuned at Ciena.

73%

signal

mean time to root-cause reduction

Watchpoint bundles incident evidence and correlates robotics telemetry automatically.

signal

hackathon tracks won on XG1

DeepLake + NomadicML wins for humanoid teleop and GR00T fine-tuning.

work_i'm_proud_of

The work is strongest when the problem is ambiguous, the constraints are real, and the system still has to ship cleanly.

field systems

signal

Shipped under hard operational constraints

Built systems that had to survive offline updates, hardware variability, noisy telemetry, and failure modes where truck rolls were the backup plan.

Proof

Offline OTA, watchdog recovery, deterministic latency validation, and edge observability workflows.

ownership

signal

Built end-to-end instead of handing off halfway

The strongest work spans architecture, implementation, validation, profiling, and the tooling needed to keep it reliable after first release.

Proof

From BSP and firmware layers to inference pipelines, replay tooling, dashboards, and deployment-safe release paths.

systems bias

signal

Production systems, not prompt demos

AI work is framed as a systems problem: retrieval quality, latency budgets, observability, rollout safety, and measurable behavior under load.

Proof

RAG evaluation loops, edge inference tuning, DPU traffic intelligence, and incident-intelligence tooling for robotics teams.

stack_map

The profile, grouped the way technical reviewers actually evaluate it.

Embedded7 tools

domain brief

Firmware and platform layers where boot, safety, and release discipline matter.

CC++Embedded LinuxYoctoBuildrootRTOSSecure Boot

Networking6 tools

domain brief

Data plane systems that care about packet paths, latency, and throughput.

P4DPDKDMA OptimizationTCP/IPBlueField DPUTelemetry

Edge AI6 tools

domain brief

GPU inference at the edge with tight latency budgets and operational constraints.

CUDATensorRTDeepStreamJetson OrinONNXNsight Systems

Robotics5 tools

domain brief

Teleop, sensing, and policy training for physical AI systems.

ROS2MuJoCoPhysical AIControl SystemsSensor Fusion

Full-Stack Systems6 tools

domain brief

User-facing control planes, APIs, and shipping infrastructure around the core systems.

Next.jsFastAPIPostgreSQLDockerCI/CDGitHub Actions

core principlehow I think

“I build systems that stay observable, recoverable, and fast under real-world constraints.”

Systems first

I care about architectures that stay observable, recoverable, and fast under real-world constraints, not just benchmark-friendly prototypes.

Ownership over handoffs

The best work usually happens when one engineer can follow the problem from vague requirement to shipped behavior and debugging evidence.

Proof over claims

I trust traces, latency budgets, rollback paths, and quantified outcomes more than polished demos or generic AI language.

why_hire_me

A concise hiring case for teams that need systems depth and shipping speed.

proof pointssystems-first

Production systems across layers

Embedded Linux, data plane networking, GPU inference, and multi-agent AI in one profile means less ramp time and fewer handoffs.

Owns ambiguity end-to-end

The work consistently spans architecture, implementation, debugging, and shipping. That matters when there is no clean spec to start from.

Built for high-ownership roles

Strong fit for teams that need one engineer to move between product, system design, and execution without losing rigor.

open_to_workavailable

Best fit

Systems-heavy roles in physical AI, robotics, edge inference, forward-deployed engineering, and embedded software.

Location

Riverside, CA

Preferred type

Physical AI / Robotics engineeringForward Deployed Engineer rolesGen AI / Multi-Agent softwareEmbedded software engineering

Open to systems, robotics, edge AI, forward-deployed, and embedded software roles

resumepdf

A compact resume for hiring teams that want the strongest signal fast. The portfolio goes deeper, but this is the shortest path to the summary.

Download resume View projects Book intro call

hire_signaldirect

Role

Systems & AI engineer

Proof

Production shipping, systems depth, and measurable reliability / performance wins

Contact

Email me

case_studies

Strongest projects, rewritten for hiring teams.

Embedded Linux0 cloud dependency

Case study

Offline OTA Update System

Field hardware needed a safe update path that worked without cloud access and never bricked the device.

Constraints

No internet dependency during delivery or install.
Updates had to prove authenticity before any state changed.
A bad release had to roll back automatically without operator intervention.

Architecture

flow

Signed bundle

Validator

Staging slot

Health checks

Promote / rollback

Tech stack

6 layers

PythonsystemdRaspberry PiCryptographyLocal HTTPShell

Outcome

Delivered an offline-first update path with atomic promotion, automatic rollback, and local audit history.

Open case study

Robotics & AI73% MTTR reduction

Case study

Watchpoint

Robotics teams needed a repeatable way to capture incidents, correlate telemetry, and stop guessing at root cause.

Constraints

Edge devices had limited resources and noisy telemetry.
The system had to preserve pre-incident context automatically.
Bundles had to be portable so engineers could replay failures elsewhere.

Architecture

flow

Incident trigger

Go edge agent

ROS2 collector

Replay bundle

RCA dashboard

Tech stack

7 layers

GoPythonROS2FastAPINext.jsPostgreSQLDocker

Outcome

Cut mean time to root cause by 73% while capturing 10K+ incidents across Jetson, Raspberry Pi, and x86 targets.

Open case study

Physical AI / Robotics2 track wins

Case study

XG1

A humanoid learning pipeline had to move from teleop to policy testing inside a 36-hour hackathon window.

Constraints

Low-latency control mattered more than raw model size.
Demonstration data had to stream fast enough to finish training in time.
The workflow had to stay useful for both teleop and diagnostics.

Architecture

flow

Quest 3 teleop

MuJoCo sim

DeepLake stream

GR00T fine-tune

Policy test

Tech stack

6 layers

Meta Quest 3MuJoCoNVIDIA SonicDeepLakeGR00TNomadic AI

Outcome

Won 2 hackathon tracks and proved the teleop-to-policy loop on the Unitree G1 within 36 hours.

Open case study

Browse the full archive

latest_blogs

Recent writing on systems, edge AI, networking, and multi-agent work.

Latest note3 min read

Agent Supervisors Should Score Evidence Loss, Not Just Answer Quality

Agent supervision often focuses on the final answer while missing a quieter failure mode: evidence gets lost, compressed, or weakened as it moves through the workflow. Strong supervisors should measure that loss directly.

AIAgentsEvaluationProduction Systems

Jul 16, 20263 min read

Read latest post

Jul 15, 20262 min read

Edge Rollouts Should Separate Device Health From Service Quality

A release can keep devices technically healthy while still degrading the service they deliver. Strong rollout evaluation separates 'the box is alive' from 'the workload is still behaving well under real conditions.'

Edge ComputingReliabilityRelease Engineering

Jul 14, 20262 min read

Robotics Debugging Gets Faster When Timing Assumptions Are Audited

A lot of robotics bugs are really timing-assumption bugs: components behave correctly only while the system stays inside a narrow implicit timing envelope. Debugging gets much easier when those assumptions are measured and reviewed explicitly.

RoboticsDebuggingTiming

Browse all writing

System Dashboard

Throughput gain@ Ciena

Hackathon wins@ 2 events

Years experience@ Industry

Active projects@ GitHub

proc monitor8 processes

PIDPROCESSYEARCATSTATUS