● Available for 2027 internships

Peyton Li

I build full-stack products and the AI that powers them.

Lately I've been building agent orchestration systems backed by knowledge graphs, RAG that verifies its own output, and NLP pipelines that detect stress in speech.

UC Berkeley/B.A. Data Science & Statistics/San Francisco ⇄ Berkeley
Email meGitHubLinkedIn

Selected work

Things I've built

The problem

Reviewing climate-action leads was slow and fragile: every pipeline failure lost context, and standing up a new agent meant rewiring the whole system.

What I built

  • Designed a 6-stage pipeline engine in TypeScript with state-driven retry, crash recovery, and approval gates, coordinating 4 agents (Sales Intelligence, CO₂ Estimation, Project Advisory, Design System).
  • Built a modular, configuration-driven agent architecture so new agents drop in through config rather than code.
  • Anchored agent decisions in a Neo4j knowledge graph spanning 50+ projects to cut hallucinations and keep every decision traceable.
  • Shipped a real-time SSE dashboard with error recovery across Neo4j and PostgreSQL so operators can review and approve in one place.

Impact

90%
faster agent integration
<30s
to approve a lead
80%
faster review turnaround
50+
projects in the graph

Stack

TypeScriptMulti-AgentNeo4jPostgreSQLSSE

Hackathons & wins

Building on the clock

Weekend sprints where I ship something real end-to-end, usually with AI at the center.

RigShare at Zero-Human Company Hackathon
Zero-Human Company Hackathon1st · Band AI + Replay QA

RigShare

Humanmade · SF · Aug 2026

A marketplace for borrowing gear at events, run entirely over text message. Lenders list the charger in their bag with a photo; borrowers text what they need and get matched to someone nearby, no app to install and no phone numbers exchanged. You pay a hold about the size of what the item costs to replace, get most of it back on return, and behind the scenes agents match borrower to lender, compare return photos against the listing, and are the only ones who can touch the money.

LinqBand AIStripePioneerTeracRender
GitHubDemoLinkedIn
VoiceSRE at Daytona HackSprint w/ Braintrust
Daytona HackSprint w/ Braintrust1st · ElevenLabs

VoiceSRE

WorkOS Office · SF · Jul 2026

An on-call AI SRE that answers the phone. It picks up Sentry and GitHub alerts, diagnoses the root cause with Llama 3.3 70B, patches and verifies the fix inside a Daytona sandbox, then calls the on-call engineer over ElevenLabs voice for approval, opening the PR while you're still on the line.

Next.jsElevenLabsDaytonaFireworks AIBraintrust
GitHubDevpostLinkedIn
VeloxEdge at AI Inference Hack Day
AI Inference Hack Day1st Place

VeloxEdge

AWS Builder Loft · SF · Jun 2026

A predictive edge cache for agentic AI. Instead of string-matching prompts (which miss ~100% of the time in agent loops), it tracks an agent's latent-space trajectory with a closed-form LinUCB bandit and pre-fetches the context, tool schemas, and memory it's about to need, driving time-to-first-token toward zero.

TypeScriptLinUCB BanditEdge CachingAgentic AI
GitHubLinkedIn
REM at JacHacks
JacHacksFinalist

REM

Founders Inc · SF · Jul 2026

Roughly 70% of serious medical errors trace back to shift-handoff failures. REM ingests nursing notes, vitals, labs, and recorded conversation into a provenance graph, writes the SBAR handoff, and runs six deterministic rules (contradiction, drift, orphaned actions, decline, falls). When one fires, it escalates up the CNA-to-physician chain by voice call with read-back confirmation, instead of waiting on a model's guess.

JacPythonDeepgramElevenLabsSignalWire
GitHubDevpostLinkedIn
SecureLoop at JetBrains Codex Hackathon
JetBrains Codex HackathonFinalist

SecureLoop

Shack 15 · SF · Apr 2026

A JetBrains-native AI security agent that closes the loop from production alert to merged fix: detect, diagnose, patch, sandbox, review, ship. It turns a real alert into an OWASP/CWE-framed diagnosis, a policy-aware patch, sandbox proof, and a developer-approved PR.

TypeScriptAI AgentSecurityOWASP
GitHubLinkedIn

The path

Where I've worked

May 2026 — Present

FlorasSoftware Engineering Intern

Agent-orchestration platform · TypeScript, Neo4j

Jan 2026 — May 2026

UC BerkeleyResearch Software Engineer

Cognitive-load & stress NLP pipeline

Sep 2025 — Dec 2025

Wadhwani AIMachine Learning Intern

Speech-alignment engine for reading assessment

Apr 2025 — Aug 2025

eNano TechnologiesSoftware Engineering Intern

SwiftUI iOS interfaces · Firebase, CoreLocation, Mapbox

Jun 2025 — Aug 2025

SF Human Rights CommissionData Analysis Senior Fellow

Program surveys, analysis, and stakeholder dashboards

The stack

What I work with

Languages & Frameworks

  • TypeScript
  • Python
  • Java
  • SQL
  • Next.js
  • React
  • Swift
  • Flask
  • Tailwind

Data & Backend

  • PostgreSQL
  • Supabase
  • Neo4j
  • Firebase
  • FastAPI
  • Express.js
  • Gemini
  • DeepSeek

AI / ML

  • PyTorch
  • SpaCy
  • NLTK
  • SciPy
  • Scikit-learn

Tooling

  • Git
  • GitHub
  • Claude
  • Codex
  • Jira
  • Playwright
  • Pandas
  • NumPy

Contact

Let's build something.

I'm looking for internships and roles where I can own features and products, iterate fast, and keep learning.

lipeytona@gmail.comGitHubLinkedIn