1. Profile
  2. Projects
  3. About
  4. Experience
  5. Contact
Yuhan Zhang
WorkAboutContact中文
Bay Area · Open to AI engineering, SWE & FDE opportunities

Yuhan Zhang

AI Agent & Full-Stack Engineer

I build AI products end to end: model and agent workflows, evaluation, reliable backends, and the interfaces people actually use.

B.S. Computer Science, Concentration in AI · University of San Francisco · Degree awarded in 2025

Authorized to work in the U.S. on F-1 OPT · no sponsorship required now
View KoalCheck LinkedIn Email Yuhan
500+
registered users
180K+
tweets in the research corpus
20K+
analyst calls tracked
270+
authors covered
Yuhan Zhang at GenAI Week Silicon Valley 2025

GenAI Week · Silicon Valley

AI Agent · Full-stack

MCP · RAG · Evals · Applied AI

GitHub Save contact Open QR code

01 / SELECTED WORK

Selected products

Two products explained by the problem they solve, how they work, and who they help.

Live · 500+ registered users

Flagship · Accountability research

KoalCheck

KoalCheck tracks stock commentators on X. When they say a stock will go up or down, it records the call and follows the result against the market—building an evidence-based track record so investors and researchers can check credibility before trusting the next call.

For AI agents

Based on that data, I also built an MCP server. AI agents can connect to access tracked X posts, commentator histories, and related real-time stock data—giving them structured, traceable evidence instead of relying only on model memory.

Research tooling only · not investment advice

How it works

  • Scores directional calls point-in-time against SPY and uses Wilson 95% confidence bounds to distinguish stronger records from thin samples
  • Serves the research data through a web product, REST, and read-only MCP tools, with SEC and FINRA sources available for cross-checking
PythonMCPRAGNext.jsTypeScriptSQLiteLLM Evals
Open live product
Phase 1 live

AI-event discovery · Edge-first community

Converge @ GPTDAO

Converge is an edge-first event-discovery and community product for AI Summit audiences. Its live Phase-1 pages and AI-event radar help people find relevant events; I took product ownership of its roadmap during my GPTDAO internship.

Live: Phase-1 public pages + AI-event radar · Auth, RSVP, and Stripe: local prototype branches

How it works

  • Shipped and maintained the public Phase-1 pages and AI-event radar on Cloudflare Workers
  • Built local prototype branches for SSR auth, concurrency-safe RSVP, and Stripe Checkout with server-authoritative pricing and idempotent settlement
Next.jsCloudflare WorkersSupabasePrismaPostgresStripe
Open Converge GenAI Week SV 2025

02 / ABOUT

Product engineering with evaluation built in

I’m a full-stack AI engineer who likes the point where product, backend systems, and evaluation meet. My flagship project, KoalCheck, delivers one research backend through MCP, REST, and a web app. My other work includes LLM-based scoring for RapBattle League and conference/community infrastructure at GPTDAO. I care about clear failure states, measurable behavior, and software that keeps working after launch.

I’m looking for AI engineering, software engineering, and applied-AI/FDE roles where I can turn ambiguous product needs into reliable working software.

USF · Computer Science

B.S. Computer Science

Concentration in Artificial Intelligence

Bay Area

California

2025

Degree awarded

AI Agent · SWE · FDE

Applied AI opportunities

03 / EXPERIENCE

Experience

Evaluation engines, conference operations, platform infrastructure, and sponsor-facing engineering.

From Jul 2025

Famiglia Doro LLC

AI Engineer Intern

Public beta · work in progress

View RapBattle League beta

A public-beta battle-rap product that turns recorded performances into transcript-based judging. Deepgram Nova-3 transcribes the audio, then rules and GPT-4o-mini score each battle against a defined rubric.

Selected contributions

  • Designed a hybrid rules + GPT-4o-mini LLM-as-judge scorer across 35 skill categories
  • Added Deepgram Nova-3 transcription, sliding-window scoring, hard-gate validation, caching, and retries
  • Shipped the FastAPI scoring service from AWS to Railway and worked across Flutter Web, Postgres, and Agora RTC
LLM-as-judgeFastAPIDeepgramFlutterPostgresAgora RTC

May–Aug 2025

GPTDAO

Full-Stack Development Intern

Built conference operations software for GenAI Week Silicon Valley and later took ownership of Converge, GPTDAO’s AI-event community platform.

Selected contributions

  • Built APIs for speakers, sessions, sponsors, and users
  • Built React admin workflows and PostgreSQL/SQLite schemas for conference operations
  • Built prototype branches for Converge authentication, RSVP, and payment workflows
Next.jsCloudflare WorkersSupabasePrismaStripe

Sep–Dec 2025

iRODS-Sponsored Project

Software Engineer

Worked with an external sponsor on data-aware compression benchmarking and client integration for research-data workflows.

Selected contributions

  • Benchmarked gzip, zlib, LZ4, zstd, and xz with repeated median and P95 methodology
  • Implemented data-aware recommendations and helped port Python baselines into the delivery stack
  • Presented project findings and participated in weekly sponsor check-ins
C++PythonCompressionBenchmarkingSponsor communication

04 / HOW I CONTRIBUTE

How I contribute

I connect AI behavior to the product and engineering systems that make it useful in the real world.

Agent products

MCP tools, retrieval, tool-using assistants, model routing, and the user experience around them.

KoalCheck · finance-report

Evaluation systems

Point-in-time scoring, Wilson intervals, LLM-as-judge, reconciliation, and explicit insufficient-data states.

KoalCheck · MacroPulse · quant-stock

Product infrastructure

Auth, payments, queues, data stores, edge runtimes, observability, and failure recovery.

Converge · KoalCheck

End-to-end delivery

Turn ambiguous requirements into working software, explain tradeoffs, and stay responsible after launch.

GPTDAO · iRODS · KoalCheck

05 / MORE SYSTEMS

More projects

Smaller labs that keep pushing the same theme: make AI outputs inspectable and testable.

Macro narrative intelligence

MacroPulse

Macro news arrives in fragments. MacroPulse brings news, economic releases, and geopolitical events into a per-asset sentiment index so researchers can see what is driving each asset in one place.

Current validation: insufficient_data · research tooling only · not investment advice

How it works

  • Scores direction, magnitude, driver theme, and affected assets per event
  • Aggregates explainable signals into a readable dashboard instead of a black-box score
  • Includes a point-in-time validation tool that returns insufficient_data when evidence is thin
PythonNLPLLM-as-scorerNext.jsEvaluation
View source

Paper-simulation validation lab

quant-stock

A local paper-trade simulator for testing what would have happened if KoalCheck calls were followed under explicit rules. It is built to audit assumptions and reconcile results—not to place real trades.

Simulation only · no real money · no proven edge

How it works

  • Versioned rules separate signal selection from fill and execution assumptions
  • Append-only event ledger with idempotent fills and nine reconciliation checks
  • 300+ automated tests across Polars pipelines and partitioned Parquet market data
PythonPolarsDuckDBParquetRule engineEvent ledger

Controlled research generation

finance-report

A reusable workflow for producing research reports without letting writing rules drift between stages. It keeps the rules in one source, re-injects them during drafting, and checks machine-verifiable constraints before delivery.

How it works

  • A single-source writing spec replaces scattered prompt rules and is re-injected at every drafting stage
  • A deterministic linter checks machine-verifiable constraints, followed by a separate style review; the workflow is packaged as a reusable Codex skill
Controlled generationSpec-as-sourceGuardrailsDeterministic QA
Ask me about it

Research or simulation tooling only · not investment advice

06 / TOOLKIT

Tools I use

Agents & AI

MCP · RAG · tool use · LLM-as-judge · evals · guardrails

Backend & data

Python · FastAPI · Node.js · Postgres · SQLite · DuckDB · Prisma

Product & infra

Next.js · TypeScript · Flutter · Cloudflare Workers · VPS · systemd

OPEN TO AI ENGINEERING, SWE & FDE CONVERSATIONS

Building an agent or applied-AI product?

If we met at AI Summit—or you’re building now—I’d be glad to compare notes. I’m exploring AI engineering, software engineering, and FDE / Applied AI opportunities.

Email Yuhan
LinkedIn GitHub Save contact QR code

Financial projects shown here are research or simulation tools. Nothing on this page is investment advice or a claim of trading performance.

Yuhan Zhang·Hosted with KoalCheck