Skip to content
View Pangqiang-Gary's full-sized avatar

Block or report Pangqiang-Gary

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Pangqiang-Gary/README.md

Gary — Applied AI/ML Engineer

I build evaluation-first AI systems: retrieval and ranking pipelines, agent workflows, time-series risk models, and production-minded ML services. I care about reproducibility, provenance, failure analysis, and making model claims testable.

Python · PyTorch · FastAPI · PostgreSQL · Docker · GitHub Actions

Selected engineering work

Project Provenance What it demonstrates Evidence
Multi-Source RAG Original build Source-specific ingestion, BM25+dense retrieval, weighted RRF, cross-encoder reranking, and explicit conflict handling Ablation report: reranking improved source-overlap hit rate from 75.0% to 91.7%
LogSentinel Original build Streaming log parsing, immutable model versions, atomic activation/rollback, tenant isolation, and online anomaly detection CI, 17 tests, 85% coverage, Loghub evaluation, Docker health check
Qlib US Walk-Forward Study Independent research workflow built on Microsoft Qlib US-equity data adaptation, explicit parameter changes, three expanding-window folds, transaction costs, and a fail-closed research gate The gate returns RESEARCH_ONLY: mean IC 0.0069, mean Rank IC -0.0053; attractive backtest returns are rejected as insufficient evidence
Native RAG Team contribution A deterministic Wiki compiler, a shared RAG Core facade, and a local FastAPI adapter in an organization repository Three merged PRs: compiler, tested Core migration, and API adapter
TradingAgents Blind Eval Original tooling around TradingAgents Forward-only evaluation controls without claiming authorship of the upstream agent system Candidate freezing, clean-commit verification, isolated memory, sealed predictions, post-cutoff scoring
Time-Series Risk Alert Original research prototype Chronological financial-ML evaluation with Transformer and classical baselines Saved datasets, model checkpoint, calibration artifacts, threshold sweeps, and v2 regime-analysis PR
MF-Structure Query Generator Team coursework; my implementation is identified in history Parsing declarative MF specifications into executable PostgreSQL query processors My generator and test-case commit, six reproducible cases, and extended-cases PR

Engineering principles

  • Evaluate before claiming. I keep metrics, ablations, failure cases, and test commands beside the code.
  • Separate evidence from prediction. Evaluation harnesses freeze inputs and provenance before outcomes are available.
  • Design for rollback and audit. Versioned state, deterministic artifacts, and explicit boundaries matter as much as model code.

Ownership and attribution

I label repositories so authorship is unambiguous:

  • Original build — I own the implementation or project design.
  • Upstream extension / research adaptation — I name the original project, my changes, their purpose, and the evidence produced.
  • Team contribution — I link merged pull requests or commits instead of presenting team work as solo authorship.
  • Learning / reference — unchanged or minimally changed reproductions, tutorials, notes, and source checkouts; these are not presented as portfolio projects.

The public collaboration evidence above comes from merged work in IDEAS-Incubator/LLM_Bootcamp_Native_RAG. A private MATS research branch is an upstream-based OpenAI Agents SDK migration, not a from-scratch system. Learning artifacts include nanoGPT Tiny Shakespeare and an unchanged OpenLLMetry fork.

Current focus

Reliable RAG · Agent evaluation · LLM observability · Quantitative and time-series ML · AI safety

Pinned Loading

  1. multi_source_rag multi_source_rag Public

    Multi-source RAG with hybrid retrieval, reciprocal-rank fusion, reranking, source weighting, and claim-level conflict detection.

    Python

  2. logsentinel logsentinel Public

    Original AIOps prototype for streaming log parsing, immutable model versioning, hot rollback, and anomaly detection.

    Python

  3. qlib-us-walk-forward qlib-us-walk-forward Public

    Research-only US equity study built on Microsoft Qlib: Alpha158/LightGBM, cost-aware walk-forward evaluation, and a fail-closed gate.

    Python

  4. tradingagents-blind-eval tradingagents-blind-eval Public

    Leakage-resistant, forward-only evaluation harness for pristine TradingAgents runs with sealed predictions.

    Python

  5. time-series-risk-alert time-series-risk-alert Public

    Original time-series risk-alert project for S&P 500, QQQ, and DJI using Transformers and classical baselines.

    Python

  6. CS562-final CS562-final Public

    Team CS562 project: my contributions implement an MF-spec parser, PostgreSQL query generator, and reproducible cases.

    Python