FRONT-END FOCUSED FULL-STACK ENGINEER · AI, MULTI-AGENT & WORKFLOW INTERFACES
GENERATIVE UI IS STILL LABELED "EXPERIMENTAL" IN MOST SDKS. I'VE BEEN RUNNING IT IN PRODUCTION SINCE 2023.
WIDGET RENDEREDNOW
STREAMPARSER · SHIPPED TO PRODUCTION, 2023
SCROLL ↓
MOST "AI PRODUCTS" ARE A CHAT BOX WEARING A COSTUME.
A model that can reason about your portfolio, your taxes, or your CRM should be able to show you the answer, not describe it in a wall of text.
CHART
TABLE
ACTION
IN 2023, THERE WAS NO PRODUCTION-READY FRAMEWORK FOR THIS.
There was no production-ready framework for this. The closest official equivalent, Vercel's streamUI, launched labeled experimental and remains paused today.
THE CATEGORY NOW //
VERCEL AI SDKCOPILOTKITASSISTANT-UIGOOGLE A2UI
WHAT HAD TO BE BUILT FROM SCRATCH INSTEAD //
01CHARTS
02TABLES & FORMS
03CRM ACTIONS
04DIAGRAMS & DOCUMENTS
themarketisup2.3%…→PLAIN TEXT ONLY. NO STRUCTURE. NO RENDER STEP.
FIG. 1: A RAW TOKEN STREAM WITH NO STRUCTURE. IT CAN SPEAK, BUT IT CAN'T RENDER.
So I built the protocol myself, and shipped it.
STREAMPARSER: A CUSTOM PROTOCOL, NOT A LIBRARY. TYPED UI, LIVE, MID-RESPONSE.
StreamParser interleaves typed JSON widget specs directly inside the LLM's token stream over WebSocket. A chart, a table, or an action card can appear while the model is still talking, not after.
TYPICAL TOOL CALLREQUEST → RESPONSE → RENDER
STREAMPARSERTOKEN → DISPATCH → RENDER, MID-STREAM
2023SHIPPED TO PRODUCTION, WHILE VERCEL'S CLOSEST EQUIVALENT (STREAMUI) IS STILL LABELED EXPERIMENTAL TODAY
30WIDGET TYPES IN PRODUCTION (THE ADVISOR PLATFORM)
20+WIDGET TYPES (AN EARLIER ITERATION OF THE SAME ARCHITECTURE, MOBILE)
1PROTOCOL, SHARED ACROSS WEB AND MOBILE, BUILT WITHOUT A FRAMEWORK
TOKEN STREAM→STREAMPARSER→TYPED DISPATCH→LIVE WIDGET RENDER
FIG. 2: TOKEN STREAM TO STREAMPARSER TO TYPED DISPATCH TO LIVE WIDGET RENDER, WITHOUT WAITING FOR A COMPLETED TOOL CALL.
SCOPE: PARSING, DISPATCH, AND RENDERING. MODEL CALLS LIVE IN A SEPARATE BACKEND SERVICE.
WATCH IT WORK.
EVERY FOLD ON THIS PAGE STREAMS THIS WAY. HERE IT IS SLOWED DOWN AND LABELED: TOKEN ARRIVES → PARSER SEGMENTS → WIDGET DISPATCHED → RENDERED.
AI AdvisorCONNECTED
How is my portfolio allocated right now?
Assistant
Pulling your holdings… done. Your portfolio has drifted 9.4% from target this quarter, driven mostly by concentrated tech gains. Here's where the allocation stands right now:
WIDGET RENDEREDNOW
PORTFOLIO ALLOCATION CHART · MID-RESPONSE
US EQUITIES
00%
INTL EQUITIES
00%
BONDS
00%
CASH
00%
Note the chart rendered while this sentence was still streaming: the widget spec was parsed straight out of the live token stream and dispatched mid-response. No completed tool call, no page reload.
SELECTED WORK //
SYS. 01PRODUCTION
B2B AI advisor workspace
NEXT.JS 16 · REACT 19 · GKE
+454%USAGE GROWTH · DEPLOYED WHITE-LABEL TO ENTERPRISE CLIENTS, INCLUDING A NATIONAL RIA PLATFORM
INTEGRATIONS //
STRIPEPLAIDSALESFORCEWEALTHBOXBLACK DIAMONDOKTA
Built the real-time AI chat renderer and design system, and wrote nearly all of the integration logic end to end, from the frontend proxy layer through the backend functions handling custodian, CRM, and payment data.
SYS. 02CONSUMER APP
Mobile AI investment copilot · iOS + Android
FLUTTER / DART
+50%ORGANIC TRAFFIC+15%RETENTION
INTEGRATIONS //
PLAIDCRYPTO EXCHANGE OAUTH
Built the earlier streaming widget architecture and portfolio aggregation that later carried into the advisor platform above.
SYS. 03EMBEDDED
Tax planning workspace embedded in a larger advisor platform
NEXT.JS · MATHJS
INTEGRATION //
TAXSTATUS (IRS TRANSCRIPT DATA)
Built a tiered strategy generator that lets advisors evaluate and re-project tax savings across dozens of strategies and time horizons instantly, instead of modeling each scenario by hand.
SYS. 04WEB + API
Professional network and marketplace platform
NEXT.JS + FASTAPI
Built the product end to end, frontend and backend: real-time messaging, a social feed, and an M&A marketplace connecting advisors to buyers and sellers.
A BIT ABOUT HOW I GOT HERE.
I started by building a consumer web app, then extended into a mobile investment copilot (still live on the App Store and Play Store today), before carrying that same streaming-widget architecture into a production B2B advisor platform, scaling it into its core AI surface.
HOW I WORK //
UI-FOCUSEDPRAGMATIC PROBLEM-SOLVERCOMPOSED UNDER PRESSURERECOGNIZED FOR FAST EXECUTION
REACT 19 · NEXT.JS · TYPESCRIPT · FLUTTER/DART · PYTHON · FASTAPI · WEBSOCKETS
I'M LOOKING FOR THE NEXT TEAM TO BUILD THIS KIND OF THING WITH.
If you're building AI products that need to do more than talk, reach out.