I'm Bill Zhang. I build LLM agent systems at Pinterest. On my own time I build voice-first projects, and I've won 36 hackathons. This page is one of the voice projects, so ask it something.
Latency, interrupted sentences, tool calls that half-fail, a model that answers the wrong question with total confidence. At RingCentral my eval pipeline moved retrieval accuracy from 90% to 95% and dropped human hand-offs from 40% to 13%. Almost all of that came out of the failure cases.
A 911 console that triages live calls, a bank you can reach from any phone, a food-rescue network that dispatches by voice, game tape you can question mid-replay. Several of them work over an ordinary phone call.
At a hackathon, I focus on getting one idea working. Dispatch AI was a phone line, a fine-tuned triage model, and a dispatcher screen, and it took the $50K grand prize at Berkeley against 900+ participants.
A phone call can be easier than navigating an app, and 1.7 billion adults have no bank account to log into in the first place. That is why TalkTuahBank answers an ordinary phone call with no app and no data plan.
★ Grand Prize, UC Berkeley AI Hackathon 2024/dispatchai↗ deep diveEmpathetic 911 dispatcher co-pilot. Aggregates emergency calls, triages severity, extracts context, and keeps humans in control.
★ Google Challenge, LA Hacks 2024/adapted↗ deep diveLectures that talk back. A voice tutor with Gemini-driven slides and emotion-aware pacing.
★ Grand + Goldman Sachs, HackUTD 2024/talktuahbank↗ deep diveVoice banking over a regular phone call. No internet, smartphone, or banking app required.
★ MLH GitHub, CruzHacks 2023/slugloop↗ deep diveReal-time UCSC loop bus tracker. Helped students see live campus bus locations and reduce pressure on metro routes.
★ NLX Overall, AI ATL 2024/soundsearch↗ deep divePhone-call voice guidance for complex websites. A caller gets step-by-step spoken help while the page highlights what to do next.
★ 2nd Place, Hack Dearborn 2024/swarmaid↗ deep diveDecentralized food rescue agents. Matches surplus suppliers to food banks and shelters with real-time logistics planning.
Every one of these was built in a weekend and defended in front of judges. $150k+ in prizes, most of them during college. Top row is the flagship and grand-prize run, bottom row the earlier and smaller ones.
Building and deploying LLM agent systems for internal and partner-facing workflows: agent design, tool use, RAG integration, and production guardrails. Also partnering with product and engineering teams to set the evaluation, quality, and cost bar each solution must clear before launch.
NowShipped a production multi-agent system automating denied-claim investigations: 3 sub-agents across 20 policy workflows, 300–500 weekly users, up to 20K claims/week. Led agent design for a major news publisher's AI agent, launched publicly in Nov 2025 over a 102-year archive in 13 languages. Owned the LLM-as-judge eval framework that took QA-audit accuracy 64% → 88% and helped close a seven-figure expansion, and cut the cost of a full health-system validation run ~4× ($4.6K → ~$1.2K).
Joined as a senior AI intern in June 2023 and converted to full-time that August. Built an LLM analytics and evaluation pipeline that grew coverage of 10,000+ weekly support chats from 1–2% to 100% while cutting processing from 8 hours to 1. Rubric-based retrieval scoring and 20 monitoring dashboards lifted accuracy 90% → 95% and dropped human hand-offs 40% → 13%.
Built an emergency-response AI platform with Berkeley SkyDeck funding, solo-engineering the low-latency voice agent and telephony behind the dispatcher demo: intent routing, emotion signals, structured handoff to responders. Dispatch AI won the $50K Grand Prize at the UC Berkeley AI Hackathon against 900+ participants.
Founded and shipped UCSC's real-time bus tracker, which served 20,000+ users. Placed Global Top 10 out of 2,000+ projects in the Google Solution Challenge.
Taught Python through a challenge-based curriculum and coached USACO students from Bronze toward Silver through algorithm practice, debugging, and contest prep.
Competed mostly during college; still enter a few events a year. Most of the projects on this page started as weekend builds.
Click a card to flip it.
Graduate-level ML, applied AI, agentic systems. Built voice-first products throughout.
Finished a 4-year program in 2.5 years. Caught the hackathon bug.
South Bay. First wrote code here.
Same information, one page, no talking required. Read it below or take the PDF; both match what the rest of this site says.
“Send me your resume.” → opens the download.
Ask about a project and the agent can look it up, describe it, and open it on the page. Retell handles the audio; a Python backend runs the agent and sends navigation events to the browser.
Browser mic stream → Retell client SDK. Low-latency WebRTC pipe to a hosted agent.
Retell bridges the audio to a FastAPI backend, where an OpenAI Agents SDK agent runs with this site's content in context and nine registered tools.
Agent emits structured tool calls: display_project(id), display_resume_page(), search_projects(query, ...), and six more.
navigation.py turns each display_* call into a navigation metadata event; the browser converts it to a bus command. No prop drilling, no router. Sections opt in.
Each section subscribes to the events it cares about and animates its own state. The site rearranges itself.
The agent says what it did while the scroll is still finishing, so the answer and the page land together.
// server: navigation.py turns a display_* tool call into
// { type: "navigation", page: "project", project_id: "dispatchai" }
// browser: metadata event → bus command → scroll
client.on("metadata", ({ metadata }) => {
const action = metaToNavigationAction(metadata);
if (!action) return;
VoiceBus.emit(action.command);
requestAnimationFrame(() => scrollToSection(action.scrollTo));
});
// every section subscribes to the same bus
VoiceBus.on((cmd) => {
if (cmd.type === "open") setOpenId(cmd.id);
if (cmd.type === "scroll") scrollToSection(cmd.id);
});