Projects

Projects

Projects

Haishi AI Agent — Manufacturing ERP Intelligence Platform

Consulting project — Solo 0→1 AI agent for a manufacturing company with real production users.

An AI-powered “Sales Copilot” that sits on top of a legacy ERP system (Zhibang ERP / SQL Server 2008 R2), enabling sales teams to query business data in natural language, get automated insights, and write back to ERP — all without replacing existing systems.

Tech Stack: FastAPI · Claude API (tool_use + extended thinking) · React 19 · TypeScript · PostgreSQL · pgvector · Docker · Huawei Cloud

Key Results:

  • 6,600+ ERP customers synced with automated nightly incremental sync
  • 25 business tables with full schema-aware AI query generation
  • Automated daily/weekly/monthly reports with AI-driven sales health diagnosis
  • Lead discovery pipeline: web scraping → LLM scoring → product-process matching (615+ companies scored)
  • ERP write-back with 2-step confirmation (preview → confirm → write)
  • Multi-tenant architecture with dual-DB isolation (Core DB + Tenant DB)
  • Admin dashboard: conversation trace, usage analytics, user feedback, sync control

Architecture: Raw → Silver → Gold three-layer data pipeline (dbt-style), proxy API for tenant isolation, FDW cross-database joins, ASP session management for legacy ERP integration.


AstrumU — ML/AI Platform

Tech Lead — Led 6-person AI/ML team building the talent intelligence platform at AstrumU (LevelSet API).

The core challenge: entity resolution at scale — deduplicating and linking people, skills, jobs, courses, and education fields across fragmented data sources (SIS, ATS, LinkedIn, military records).

Tech Stack: Ray Serve · vLLM · Kafka CDC · pgvector HNSW · Dagster · FastAPI · PostgreSQL · MLflow

Key Results:

  • Multi-model serving: 6+ Ray Serve deployments with scale-to-zero, 10K+ entities/day online inference
  • Fine-tuned Qwen-2.5 (1.5B-7B) with LoRA SFT for structured extraction — 1-2 orders of magnitude cheaper than API
  • Multi-stage RAG pipeline: retrieve → generate → retrieve again → rerank (Google Search + pgvector dual retrieval)
  • Event-driven pipelines with Kafka CDC + SCD patterns, replacing Databricks batch jobs
  • Unified PostgreSQL platform replacing fragmented Databricks/Postgres/feature store architecture

Deep Dive: The Evolution of MLOps: A Six-Year Journey at AstrumU


CityDecoded — AI-Powered Zoning Copilot

Side project — SaaS platform making zoning codes and municipal regulations accessible through AI.

Property owners, developers, and homeowners often can’t answer basic questions about what they can build — because zoning codes are written in legal language, buried across hundreds of pages, and vary by city. CityDecoded makes this information actionable.

Tech Stack: Next.js · Supabase (Postgres + Auth + Edge Functions) · Google Gemini · MCP · Airflow · Tailwind CSS

Key Features:

  • Dual-window interface: code browsing + AI chat side by side
  • Property-specific reports: buildable envelope, expansion potential, permit roadmap
  • Developer API for third-party integration
  • ETL pipeline for ingesting and structuring municipal code documents

Writing

Agentic Data Engineering Series

A 9-part blog series on LinkedIn exploring the data infrastructure requirements for production AI agents. Core thesis: your AI agent’s ceiling isn’t the model — it’s the data underneath.

Topics covered: data quality as agent reliability, semantic layers, context engineering, agent-as-teammate patterns, AI-native team structure, and why you can’t vibe-code a data system.

Read on LinkedIn →

Selected Blog Posts