// Software Engineer · AI agents & streaming systems

I build AI agents and the systems that prove streaming works.

SDE II at Amazon Prime Video. I measure playback quality on hundreds of real devices, and I build the platforms and LLM agents that run, debug and fix that lab.

See the work GitHub Get in touch
~270
physical streaming devices automated
63
tools in a production LLM agent
8
playback QoE metrics independently measured
95%
automation across 3,688 regression cells

Selected work

Architecture at a public level of detail. Demo videos coming.

Streaming deviceFire TV · Roku · Apple TVNetwork switchimpairment by policingClient telemetrywhat the app reportsHDMI capturelive video streamVision + timingframes → eventsMeasured QoETTFF · rebuffer · bitrateComparisoncause per disagreementreported
Video measurement · QoE

An independent witness for playback quality

A streaming client reports its own startup time, so checking it against its own telemetry proves nothing. I built a measurement system that watches the actual video output of each device and times what the customer sees.

  • Measures time-to-first-frame, rebuffering, bitrate, bandwidth, abandons and fatal errors on every device family
  • Network impairment via switch-port policing to induce real-world conditions
  • Showed a fleet-wide startup metric understated what customers experience by 1-3 seconds
PythonVideo capture (SRT)Vision modelsAWS
Engineerasks a questionAgent loopLLM on BedrockRead-only toolsresults · logs · telemetryRun toolsrun test · watch live videoRemediationdiagnose → proposeTyped confirmation→ execute → audit → verifyhuman approves high-risk changes
LLM agents · Tool use

A lab agent that runs tests, watches devices and proposes fixes

A tool-use agent on AWS Bedrock with 63 tools: it can run a test, read a device's live video and telemetry, and diagnose a failure. Fixes go through a gated path: diagnose, propose, typed confirmation, execute, audit, verify.

  • Risk graded by what a change touches, with rate limits counted from the audit log
  • Smaller models for mechanical steps, larger ones only where they pay off
  • Designed the tool-calling architecture of the team's support bot, built with an intern I mentored
AWS BedrockAgent designTypeScriptPython
Web appReact · 25 routesSchedulercron jobs · device groupsAPI + LambdasPython · API GatewayLive videoin-browser device viewLab monitorpolls the lab every 10 minState storeDynamoDB · alarmsWarehouseRedshift stats · reports
Full-stack platform

One operations surface for a 270-device lab

The team had four separate tools for scheduling, accounts, debugging and device state, and none of them belonged to it. I built a single platform for all of it, with no spec and nobody asking for it, and it became the team's daily driver.

  • Job scheduler, live device state, in-browser live video of any device, statistics and reports
  • Lab monitor raised reachable devices from 4 to 37 and cut open faults from 39 to 6
  • 111 React components, 8 CDK stacks; authored 82% of 1,400+ commits
ReactTypeScriptLambdaDynamoDBRedshiftCDK
Coordinator agentsees every threadAgent 1own workspaceAgent 2own workspaceAgent 3own workspaceAgent 4own workspaceShared knowledge basefacts · gotchas · decisionsDashboard + MCP toolslive state · overlap warningsharvest
AI-native engineering

A multi-agent workflow for parallel engineering

I run many investigations at once, each as its own agent with its own workspace. A coordinating agent watches all of them, and a shared knowledge base keeps what each one learns, so later work starts from what earlier work found.

  • Agents from more than one model family share the same tools, skills and knowledge
  • Dashboard and MCP tools show live state and warn when two agents edit the same file
  • Phone access and notifications when an agent finishes or needs a decision
Multi-agentMCPKiroCodex
Test playbackads + live eventsPlayback datasessions · eventsAd + audience datalate-arriving rowsNightly validationfield-by-field checksNightly reportto data ownersFindingscause per mismatch
Ads · Audience measurement

Validating ads and audience measurement end to end

Audience measurement is reported to outside partners, so mistakes in it cost money. I automated validation of audience-measurement and ad-insertion telemetry across several datasets that update late, for an ads team and a program team at once.

  • Nightly checks of every required telemetry field on every device family
  • A cause attached to every disagreement, rather than a pass/fail per night
Ad techData qualitySQLJava

Skills

What I reach for.

AI & LLM

Tool-use agents on Bedrock, multi-agent workflows, MCP, vision models in production pipelines, prompt and tool design.

Streaming & media

Playback QoE (TTFF, rebuffer, bitrate), SRT and HLS video, ad insertion, audience measurement, Fire TV, Roku, Apple TV and Android.

Cloud & backend

AWS CDK, Lambda, DynamoDB, Redshift, Athena, S3, EventBridge, API Gateway, EC2, ECS; event-driven, fault-tolerant design.

Languages & frontend

TypeScript, Python, Java, Swift, Go, C/C++, SQL; React.

Experience

5 years across streaming, defense, research and government systems.

2025 - now

Software Development Engineer II · Amazon Prime Video

Playback quality measurement, lab platform and AI agents. Promoted from SDE I.

2024 - 2025

Software Developer · Inalab Consulting

High-volume data pipelines and monitoring for DoD systems.

2023

Application Engineer · Sandia National Labs

Streaming sensor-data pipelines for DoE space data acquisition.

2021 - 2023

Software Developer · NMT ICASA

Spring Boot microservices and React for an NM Department of Justice case-management system.

2022

B.S. Computer Science · New Mexico Tech