Capabilities

Turning experience into assessable capabilities

This page is organized by the problems I can solve rather than by a resume timeline. Each capability maps back to concrete work so it can be inspected.

Agent Harness / task execution systems

Putting models inside controlled task systems

The core concern is not a single model answer, but how requirements, context, tools, state, outputs, logs, and failure recovery become a stable engineering loop.

Problems handled

How long-running tasks are queued, executed, paused, failed, and resumed.
How models modify files, generate reports, or call business tools within limited tool boundaries.
How each run leaves timeline, artifact, snapshot, log, and replay material.

Related work

GraphPilot, Hovo, and Guanlan.

GraphRAG / knowledge assets

Turning documents, examples, and experience into retrievable knowledge

Rather than relying only on vector search, I combine document structure, entity relations, context expansion, rerank, and fact constraints to support agents and business Q&A.

Problems handled

How enterprise documents become stable document trees and knowledge nodes.
How examples, APIs, rules, and historical errors become usable agent assets.
How retrieval results can be audited, debugged, and improved.

Graph application engineering / complex relationship data

Turning relationship data into business operation interfaces

I have long worked with knowledge graphs, organization structures, equity relations, data lineage, workflow dependencies, and business topology, focusing on display, interaction, editing, and system embedding.

Problems handled

How relationship data is modeled as nodes, lines, business data, and runtime state.
How multiple framework adapters share a core engine instead of duplicating logic.
How graph editing, minimap, export, layout, and events become reusable engineering capabilities.

Financial data systems / banking semantics

Understanding data constraints in financial business systems

Earlier experience with bank customers, enterprise relationships, risk alerts, external data governance, and knowledge graph scenarios shaped later AI system design around auditability, evidence chains, and data boundaries.

Problems handled

How business objects, metric definitions, attachments, and human review form task context.
How model output is constrained and reviewed under strong rules and audit requirements.
How financial data analysis systems preserve evidence between queries, results, and reports.

AI data analysis / local research systems

Breaking analysis into reviewable steps

Whether in banking analysis or local quantitative research, the key is not one-shot conclusions, but preserving the relationships between data processing, queries, metrics, charts, reports, and replay.

Problems handled

How natural-language analysis becomes constrained SQL, structured outlines, and interactive reports.
How local data forms indexes, relation caches, trigger candidates, and replay records.
How metric output states its boundary and avoids treating backtest statistics as certainty.

Productized delivery / small-team system execution

From core code to usable products

My experience covers requirement breakdown, architecture design, core development, docs and examples, deployment, user feedback, and continued iteration, especially for complex tool products from zero to one.

Problems handled

How to package low-level capability into documents, examples, and tool entry points that developers can understand.
How to build logging, diagnosis, feedback, and iteration channels from real user tasks.
How open-source projects, knowledge assets, online agents, and enterprise services support each other.

AI system governance / auditable execution

Designing auditable AI execution systems

Permission boundaries, tool calls, SQL safety, runtime logs, evidence chains, task snapshots, and audit records must be part of the system design so AI execution can be reviewed and governed.

Problems handled

How to restrict model access to files, databases, and business tools.
How generated results are automatically bound to inputs, tool calls, datasets, and artifact records.
How errors or disputes can be traced back through the full execution chain.

Related work

GraphPilot, Hovo, and Guanlan.

Scenario application platform / app packages and workpapers

Letting different business scenarios enter one AI platform

Scenario apps can be organized as app packages, templates, tools, node workflows, runtime configuration, and workpapers, while the platform unifies execution, audit, and release governance.

Problems handled

How different business scenarios own their templates, tools, nodes, and workpaper structures.
How the platform manages app release, version validation, runtime config, and task state.
How scenario apps expose OpenAPI or other entry points without bypassing governance.

Related work

Hovo and GraphPilot.

Agent evaluation / replay systems

Turning real tasks into samples for continuous improvement

Task inputs, context, tool calls, errors, repairs, artifacts, and business results are saved for replay, model comparison, system diagnosis, and knowledge asset iteration.

Problems handled

How to build repeatable evaluation samples from real agent tasks.
How to compare results after changing models, prompts, tool strategies, or knowledge packs.
How failure cases become inputs for the next system improvement cycle.

Business semantics / controlled data access

Wrapping raw data as governable business semantics

Table structures, metric definitions, object relations, query permissions, and tool calls can be wrapped into a business semantic layer so AI and applications access data through controlled operations.

Problems handled

How database tables, business objects, and metric definitions become understandable semantic objects.
How models fetch data only through controlled operations or safe SQL.
How queries, results, reports, and business explanations remain linked.

Related work

Hovo and Guanlan.

Complex tool products / developer experience

Making complex capabilities reliable for developers

API design, example systems, documentation structure, error messages, knowledge assets, and AI-assisted paths work together to make complex tools usable and hard to misuse.

Problems handled

How stable APIs, examples, and docs lower the barrier for complex graph components.
How knowledge assets and examples reduce API misuse in AI-generated code.
How open-source feedback becomes documentation, examples, and product capability.

Related work

relation-graph and GraphPilot.

Data-intensive systems / local research and experiments

Building data research systems that can evolve over time

Local data warehouses, indexes, caches, offline computation, lightweight online reads, replay records, and experiment iteration need clear layers to stay reproducible.

Problems handled

How historical data, working indexes, relation caches, and replay results are stored in layers.
How heavy recomputation is moved offline while online usage stays lightweight.
How strategy, retrieval, or model experiments leave comparable and reviewable records.