Skip to experience
Qingping Champ
中文Let's connect

ENGINEERING THROUGH-LINE

Building agents on reliable data systems

I design adaptive vector systems, traceable omni-modal retrieval pipelines, and recoverable agent runtimes with explicit evaluation gates.
Creative desk with a computer, notes, coffee, and a small robot
Evidence, retrieval, and runtime decisions on the same desk
OPERATING THESIS

Reliability is a through-line

Reliable Agent ExecutionOmni-modal RetrievalAdaptive Data Systems

From databases and retrieval to agents, one thread runs through my work: making complex systems understandable, useful, and trustworthy.

Engineering experience

Selected responsibilities and technical decisions.

AI Agent Engineering

ByteDance

AI Agent Algorithm Engineer

Designing traceable agent engineering workflows with staged context, permission and evidence gates, and evaluation-driven rollout.

  • Built a layered evidence system for business knowledge and code, then compiled task-specific context by development stage.
  • Connected approvals, CI, deployment, and human review into an auditable loop from requirement to verification.
  • Co-designed a tool-grounded multi-agent review flow with threshold-based adoption and rollback.
AI AgentEvidence GatesEvaluationCI Workflows
Analytical Databases

ByteDance

Lakehouse and Retrieval Backend Engineer

Delivered multi-source knowledge ingestion, hybrid retrieval, and multimodal indexing inside a database team, connecting data engineering, vector search, and production deployment.

  • Completed an end-to-end knowledge pipeline for authentication, parsing, chunking, embedding, storage, hybrid recall, and reranking.
  • Validated layered scene and shot models for audio and video with multi-channel recall, RRF, and rerankers.
  • Contributed to vector-engine capabilities, system-table extensions, memory-store integration, and DiskANN indexing.
ByteHouseClickHouseHybrid SearchDiskANN
AI Infrastructure

JD Retail

AI Infrastructure Software Engineer

Built shared retrieval foundations and intent-driven recommendation workflows for enterprise knowledge and merchant services.

  • Designed a multi-source hybrid retrieval framework across semantic vector search, keyword search, and intent routing.
  • Led an end-to-end Doc2Query pipeline that addressed the semantic gap between user language and platform knowledge.
  • Contributed to a recommendation flow combining user profiles, metric evaluation, action plans, and service matching.
VearchElasticsearchDoc2QueryWorkflow
RESEARCH THREAD ·2024 to present

One systems thread across data, retrieval, and agents

My work spans database systems, adaptive indexing, multimodal retrieval, and agent runtimes. Across that sequence, the emphasis shifts from measurable adaptation to traceable evidence and recoverable execution.

  1. 01Database systems
  2. 02Adaptive vector indexing
  3. 03Retrieval infrastructure
  4. 04Reliable agent execution

Systems outside the job title

Three independent build threads, shown here as a progression rather than a second resume.

Public-source project

Tessmora

A self-hostable omni-modal retrieval platform where documents, images, audio, and video keep their semantic units and join one traceable answer pipeline.

Read projectGitHub
  1. Uses source-faithful agentic chunks for documents, VLM and CLIP for images, ASR and CLAP for audio, and scene-shot-key-frame structures for video.
  2. Combines dense, sparse, visual, audio, and video retrieval through weighted RRF and Cross-Encoder reranking.
  3. Runs complementary Agent queries through the same read-only retriever and streams type-specific citations for documents and media.
Independent design and development

Napier

A local-first glass-box agent runtime that records messages, model calls, tools, goals, branches, and artifacts from long-running tasks in one ordered evidence ledger.

Read projectGitHub
  1. Uses a transactional SQLite WAL ledger with strictly increasing per-thread event sequences; chat is one projection.
  2. Supports explicit resume, sequence-accurate branching, hash-verified replay snapshots, independent no-tool evaluation, and inspection across Web and CLI.
  3. Checks workspace and tool policy immediately before execution and records control decisions in the same evidence model.
Research system

Adaptive Vector Index

A query-adaptive ANN index that combines recall-targeted scanning, online updates, and cost-based partition maintenance.

Read project
  1. Uses a rolling hit window and measured scan-latency grid to identify overloaded and underused partitions.
  2. Adjusts partition scanning around a per-query recall target instead of assigning every query a fixed nprobe budget.
  3. Combines online split and reassignment with a C++ compute path that supports batched scanning, AVX-512, and NUMA locality.
CAPABILITY MAP

What connects the work

Execution, evidence, and data structures are treated as one system, not three disconnected specialties.

01

Intelligent systems

Reliable agent executionEvaluation and rollbackPermission and evidence gatesWorkflow automation
02

Retrieval and data

Hybrid retrievalOmni-modal evidenceAdaptive vector indexingKnowledge ingestion
03

Engineering

TypeScriptPythonC++ReactBackend and infrastructure
SEE THE WORK

The project pages show how those decisions become systems.

View projects