Software architect · AI-systems builder · Investor

I build AI that writes01, narrates02, and invests03.

Three production pipelines — each shipping a finished, rigorous, real artifact: a writing editor that rebuilds voice, a storybook engine that narrates and illustrates, and an investment system that enforces discipline. Former engineering director who led 50+ engineers; full-stack to multi-agent AI.

Work with me See the work No slideware — open a proof artifact
3
Production
pipelines
45 detectors
EN + ZH voice
patterns stripped
37 rules
Encoded in the
investing engine
2 langs
Everything ships
EN & 中文
§ 01–03

Selected work

Each system ships a proof artifact you can open — not a pitch.

01Narrated storybook engine

Fairytale Engine

An end-to-end pipeline that turns a story idea into a narrated, illustrated children's book — script, voice, illustration, mixed audio.

Pipeline
Script → voice → illustration → mixed audio
Continuity
Per-character visual manifests — no drift across plates
Output
KDP-ready EPUB · narrated read-along
Hear it & see the book
Plate · illustration Rocco & the Trail to the Top
Watercolor illustration: three raccoon siblings hiking a forest trail toward a mountain, from the picture book 'Rocco & the Trail to the Top'.
Rocco · Benny · DottyPlate 08
02Decision-support system

Investment Pipeline

A staged decision engine that encodes a personal investing discipline — macro → technicals → valuation → risk — into a structured, redaction-verified report.

Gates
7 staged gates · 37 encoded decision rules
Discipline
Backtested thresholds · versioned releases
Privacy
Public report only — no holdings, sizing or P&L
Read a redacted report
Public report render_report.py v4.121
NFLX Sample · structure only
WATCH
Macro regimeGREEN
Support / resistance75.01 — 108.95
Gamma regimeCUSHIONED
Valuation · sizing · targets▮▮ redacted
Redacted · no holdings · not financial advice
03Bilingual writing editor

Editor Pipeline

A 7-stage editor that strips AI-generated patterns and rebuilds a writer's real voice — in English and 中文.

Detectors
24 EN  +  21 ZH pattern detectors
Model
Per-writer voice profile, fitted from a corpus
Verified by
Test-driven eval harness · cross-model judge
See the before → after
Voice diff · sample EN
AI draft

In today's fast-paced world, it is crucial to delve into the multifaceted tapestry of one's thoughts.

Li's voice — rebuilt

I keep one true sentence and cut the rest. The reader feels the difference before they can name it.

§ 04

How I work

One belief, applied three ways.

Engineer who writes. Investor who builds.

The three projects aren't unrelated demos — they're one belief, applied three ways: AI is most useful when it carries judgment and craft all the way to a finished artifact.

More about me
A

Proof over pitch

Every project ships an artifact you can open — a voice diff, a narrated book, a redacted report — instead of a deck about one.

B

Verify ruthlessly

Prompts are treated like software: fixed corpora, deterministic checks, and an independent cross-model judge. "It works" is measured, not asserted.

C

Judgment as a system

A value-investor's discipline applied to engineering — encode the rules, stage the gates, and let the structured output carry the call.

Let's build
something real.

Advisory on AI architecture and multi-agent systems — or just to compare notes. No hard sell.