AI Vibe Coding Projects
These projects are where I turn finance, strategy, and analytics work into reviewable AI systems. My focus is not one-off prompting; it is building workflows with grounded retrieval, typed artifacts, human review checkpoints, reproducible outputs, and deployable interfaces.
Strategy + finance judgment
DCF, market-entry diligence, customer value-chain analysis, and investment memo logic.
Grounded AI systems
RAG, cited evidence stores, source manifests, provenance files, and refusal gates.
Structured artifacts
Pydantic/Zod schemas, validation, review handoffs, tests, and reproducible reports.
Product shipping
Next.js, Streamlit, Quarto, GitHub Pages, Vercel, and recruiter-friendly demos.
AI DCF Analyst Coworker
Live demo
A source-grounded investment research coworker that turns public-company evidence into reviewable DCF valuation artifacts.
What I built
- Runtime SEC Company Facts and 10-K evidence ingestion.
- Typed LLM outputs for company profile, competitive landscape, synergies, forecasts, WACC, and valuation.
- Human review gates before AI artifacts can feed downstream valuation steps.
Evidence
- Next.js 16, React 19, TypeScript, Zod, Vercel.
- 81 passing tests covering schemas, ingestion, forecast, WACC, and valuation logic.
Capability signal: I can translate investment-analysis judgment into a controlled AI workflow where sources, assumptions, and model outputs remain inspectable.
Decoupling Analyst
Active build
A local-first AI workflow for Teixeira-style customer value-chain analysis, grounded research, bilingual strategy reports, and decoupling recommendations.
What I built
- Tavily-grounded retrieval plus RAG over Teixeira course methodology.
- 14 Pydantic-typed modules from company profile to weak-link diagnosis, decoupling thesis, business model, competitive response, and critic.
- English and Chinese reports with evidence IDs, provenance, cost summaries, and a read-only Streamlit showcase.
Evidence
- Grounding gate refuses to ship source-free reports when no visited URLs are retrieved.
- Teixeira calibration: 57% exact match, 81% exact-or-partial match, about $0.30 per grounded run.
Capability signal: I can encode a strategy framework into an auditable AI pipeline, then evaluate it against classroom cases instead of trusting polished prose.
Quant Finance Lab
Research lab
An A-share ETF rotation research lab for turning technical-factor ideas into reproducible signals, backtests, diagnostics, and model baselines.
What I built
- ETF universe setup, AKShare data download path, daily CSV ingestion, and factor calculation.
- Rotation backtests with position sizing, fees, slippage, stop loss, take profit, trailing stop, and max-position constraints.
- Walk-forward logistic regression baseline that records
model_up_probabilitywhile checkingmodel_last_label_date < date.
Evidence
- Produces metrics, trades, signals, equity curves, monthly returns, parameter sweeps, article-gap analysis, and HTML reports.
- Separates reproducible local research from future QMT simulation or live trading stages.
Capability signal: I can build finance experiments that make assumptions testable, expose leakage risk, and leave audit-ready outputs behind.
AustinLiWeb
Portfolio system
A Quarto portfolio and publishing system that turns my resume, analytics blogs, and AI project work into a public professional presence.
What I built
- A GitHub Pages site with separate resume, AI projects, and analytics blog surfaces.
- PDF resume embedding, project taxonomy, public links, and generated
docs/output for deployment. - Iterative content/design workflow that keeps recruiters focused on real projects instead of scattered repo links.
Evidence
- Built with Quarto, custom CSS, tracked published output, and repeatable local rendering.
- Packages technical work from finance, analytics, and AI into a navigable portfolio artifact.
Capability signal: I can ship and maintain the presentation layer around technical work, which matters when analysis needs to be understood by non-engineering audiences.