jamesli.work

I build AI systems for evidence-backed decisions

Own products, client delivery, and open-source systems across recruitment, tendering, commerce, search, and evidence.

AI-native systems builder · Brisbane / Remote

View selected work Read resume
Illustrated portrait of James Li
Selected work

Selected work

Six systems. Clear ownership. Honest limits.

01OptiTalentOwn ProductEvidence-first recruiting workflowsOpen case Glass recruiting workspace with evidence profiles and reviewed shortlist states02Tender LubanClient DeliveryTender review and commercial bid assemblyOpen case Tender qualification evidence trays leading to a reviewed commercial document03KinmedClient DeliveryB2B medical web and programmatic SEOOpen case Clinical product discovery surface with medical material and global market routes04MuriersilkClient DeliveryD2C commerce and SEO recoveryOpen case Silk commerce still life with a transparent product surface and growth path05MemtruthOpen SourceEvidence and document runtimeOpen case Evidence instrument connecting source documents to a central traceable claim06PiveoOpen SourceSelf-hostable AI creative workstationOpen case Creative workstation with media gallery, generation queue, and comparison surface

Recall is a system, not a vector-search feature.

I design retrieval from exact and lexical recall through hybrid search, evidence-preserving fusion, learned-ranking experiments, and local compute placement. Models ship only when the evaluation earns it.

  1. Exact + TF-IDF

    Exact, phrase, trigram, full-text, and character n-gram recall create a deterministic lexical foundation.

    • Exact match
    • BM25
    • TF-IDF
    • Full text

  2. Vespa + Qdrant hybrid

    Dense, lexical, normalized-exact, full-resume, and evidence lanes broaden the candidate pool across hosted and local runtimes.

    • Vespa
    • Qdrant
    • Dense recall
    • Evidence lanes

  3. Evidence-preserving fusion

    Weighted reciprocal-rank fusion and per-lane reserves keep candidate inclusion inspectable instead of collapsing it into one opaque score.

    • Weighted RRF
    • Lane reserves
    • Provider trace

  4. LTR evaluation

    evaluated

    Learned rankers are measured against deterministic baselines and only advance when the evaluation supports the added complexity.

    • LightGBM
    • LambdaRank
    • Frozen protocols

  5. Local compute placement

    Local embedding, tenant-scoped Qdrant recall, and Rust-owned fusion move appropriate work closer to the data and user.

    • Local embedding
    • Rust
    • Tauri
    • Scoped Qdrant

How I work

AI writes most of my implementation. I design the capabilities, constraints, interfaces, and evidence that make it reliable.

Product path

  1. Business domain core

  2. Application capabilities

  3. CLI + MCP + API

    TUITUI is a development-stage observability sidecar.

  4. Thin GUI validation

  5. Complete product GUI

Frontend path

  1. Business goal

  2. IA / site tree

  3. Design Read

  4. Desktop + mobile references

  5. Image-to-Code

  6. Browser verification

  7. Shipped page

Domain experience

Where I am useful.

I spent years operating real businesses before building software for them. That makes me unusually attentive to workflow, adoption, risk, and the difference between a demo and a useful system.

Brisbane / Remote across APAC

Agent-native operational systems

Capabilities shared by people and agents, with clear state, review, risk, and audit boundaries.

Search, retrieval, and ranking

Recall, fusion, ranking, evidence, and compute placement designed as one product system.

Evidence and document systems

Source-aware extraction, provenance, citations, structured outputs, and reviewable decisions.

Commerce, growth, and SEO

Information architecture, programmatic discovery, structured data, and commerce-safe delivery.

Product infrastructure and delivery

Typed contracts, multi-tenant boundaries, release gates, deployment, and live verification.

Skills & Stack

Concrete tools and working strengths, grouped by the systems they help me ship.

Working strengths

  • Systems thinking
  • Workflow discovery
  • Requirement translation
  • Evidence-led judgment
  • End-to-end ownership
  • Human-agent interaction

Core Languages

  • TypeScript
  • Rust
  • Python
  • Java
  • SQL

Frontend

  • React
  • React Router
  • Next.js
  • Astro
  • Vite

Backend & Interfaces

  • Node.js
  • Hono
  • FastAPI
  • Django
  • REST
  • MCP

Retrieval & Data

  • PostgreSQL
  • pgvector
  • Vespa
  • Qdrant
  • BM25
  • TF-IDF
  • RRF

Ranking & AI

  • LightGBMevaluated
  • ColBERT
  • Embeddings
  • Structured outputs
  • Evaluation harnesses

Infrastructure

  • Cloudflare
  • Docker
  • GitHub Actions
  • Dokploy

Product & Web

  • Information architecture
  • Site trees
  • Technical SEO
  • Schema.org
  • Imagegen
  • Image-to-Code

Open source

Public systems and framework contributions that survive review.

Claim discipline

Subsystem ownership is earned through review and maintenance.

I label ownership and contribution state precisely, and keep framework work distinct from the central runtime.

Education & Contact

Operating context, formal study, and a direct path to the next useful conversation.

2025Master of Science in FinanceAsset ManagementBrest Business School
2021Diploma of BusinessBusiness AdministrationTAFE Queensland

LanguagesEnglish · Mandarin · Cantonese

Brisbane · Remote across APAC

Have a hard workflow worth making useful?

I am open to AI solutions, forward-deployed, product, retrieval, and implementation work.

Start a conversation Read the resume