Yen-Nan (Aaron) Ho

Data Scientist, Harbor Tech | Taipei, Taiwan | Oct 2025–Present

Worked on two Harbor AI 2.0 products sharing a technical foundation: ongoing equipment monitoring/diagnostics and unit-by-unit end-of-line quality inspection. I lead product/domain and customer-delivery work; AI agents and engineering partners contribute implementation and verification as described below.

Industrial Intelligent Monitoring Platform: From Domain Workflows to an Independent Product

• Product and operator workflows: Led product/domain design and built much of 2.0's functionality from scratch with the team. Translated vibration, PRPD, RUL, trends, and data quality into layered equipment/sensor status, event views, and operator workflows. Used a device-first reporting flow and connected multilingual interfaces, charts, mobile layouts, and live APIs.
• Legacy-system evolution: Defined ownership of equipment data versus user settings during coexistence, protecting new settings from legacy synchronization. Advanced device-control transfer, event-source replacement, and auditable migration; stopped imports when sensor mappings were unclear and revalidated after corrections. Confirmed in October 2026 that 2.0 operates independently of 1.0; this does not imply every customer or dataset has migrated.
• Analysis workflows: Collaborated with platform engineers on Django orchestration, defining APIs, data access, artifact-by-reference contracts, task lifecycles, and failure behavior. Integrated DB-authoritative per-sensor scheduling, manual triggers, batches, and history, choosing aggregate or WAV inputs by availability.
• Edge data and control: Owned Gateway ingestion specifications and design validation across uploads, heartbeats, duplication, integrity, and capacity, handing server implementation to the responsible engineer. Also contributed server-side software-package distribution and RUL-to-threshold feedback across platform state/history and Gateway updates, keeping ingestion, migration, OTA, and threshold policies distinct.

Equipment End-of-Line Quality Inspection Platform: Requirements, Operation, and On-Site Delivery

• End-to-end responsibility: Led customer requirements, specifications, implementation, UI/UX improvement, and on-site delivery with agents and engineering partners. Defined unit/batch relationships, reference samples, multipoint measurement, interrupted orders, serial/work-order continuity, human review, and reports. Kept functional verification separate from on-site decision-threshold calibration.
• Operator experience: The initial AI-generated UI met specifications but required five to six page levels and circular navigation. Reorganized daily work into plan-list, work-order, and board interfaces, integrating reports, review, and reference management or moving them into dialogs. Added focused primary actions, arrow/Enter controls, status cues, and sensor-disconnection reminders. DQA feedback was qualitative.
• Measurement safeguards: Separated sensor-installation confirmation from measurement, using sensor IDs and installation diagrams as mandatory pre-measurement checks.
• On-site judgment and fixes: Installed/demonstrated the system, trained users, and answered customer questions on October 12, 2026. Personally diagnosed a clock offset behind interrupted connections, executed the troubleshooting commands, and completed on-site and follow-up fixes. Claude supported discussion of symptoms and hypotheses. Clarified feedback for the team and explained operation and design choices to users.
• Delivery status: Coordinated DQA, test versions, manuals, and acceptance records. The project remains ongoing; formal sign-off and final acceptance remain unconfirmed. My estimate of roughly 80% overall contribution covers the end-to-end work, not a measured code-authorship share.

Models, Diagnostics, and Reports: Keeping Conclusions Tied to Data

• Assumption review: Traced RUL analysis to test constants treated as observations, rejected the conclusion, and required multi-parameter/multi-curve and forward/reverse ablations. Connected research notebooks to product code to check rapid-degradation misses, healthy-series false alarms, and implementation drift.
• Diagnostic/output rules: Connected preprocessing, valid operating segments, fault classification, ISO severity, RUL, and reporting while keeping fault type, severity, and remaining time distinct. Reviewed sampling, coverage, selection bias, and denominators; explicitly degraded unavailable results and preserved engine outputs. Field lifetime ground truth and confidence calibration remain incomplete, so I do not claim RUL accuracy.
• Report-service productization: Turned manually validated field reports into reusable methods and quality rules. Used a July LangChain/LangGraph POC to validate node contracts, device isolation, partial regeneration, and fallbacks. Deterministic nodes produced values/classes; LLMs organized narrative. Advanced service/platform integration in August and mainline DOCX/frontend downloads in September. Led methods and quality criteria with team/agent implementation and integration; deployment and customer acceptance are tracked per path.
• Model-service interfaces: Validated live vLLM chat, embedding, and tool-calling capabilities, separating connectivity, functional, and semantic health checks. Endpoint availability is not evidence that every report path has switched providers.

AI-Assisted Engineering and Release Delivery: Making Parallel Work Traceable

• Established shared requirements, analysis, decisions, issues, tests, and version records. Decomposed work, supplied persistent context/tools, and set verification/stop criteria; agents researched, implemented, repaired, and rechecked changes on isolated branches.
• Ran daily functional regression and monitoring, routing findings to issues and Bitbucket PRs. Reviewed data assumptions and product behavior with human reviewers before merge/release. Fed recurring problems into reusable rules and tests; no overall productivity multiplier or automatic-repair success rate has been measured.
• Integrated Docker/offline packages, cross-architecture installation, version information, and platform update/rollback. Separated update execution from platform downtime risk, with prechecks, progress, failure signals, and recovery. Used real-machine failures to improve write freezes, schema/constraint, storage, and restart checks.
• Refined delivery-package contents, excluded customer raw data, validated clean installations, and led release integration/readiness decisions. Migration checks covered fields, time zones, readability, and continuity; linked code, issues, versions, and manuals through VitePress.

Field Analysis, Technical Delivery, and Collaboration

• Aligned field/quality teams on DB, event, and WAV scope for semiconductor and large multi-sensor partial-discharge analysis. Used the full database to locate events, then retrieved corresponding waveforms for batch analysis, accumulated PRPD images, AI drafts, and page-by-page human review.
• Used max/P99 to distinguish sustained motor/pump severity from one-second spikes, checking operating segments, RUL basis, and VRMS variability. Later used minute medians, nonparametric tests, multiple-comparison correction, and effect sizes to reassess customer observations. Delivered the analysis and customer explanation; root cause still requires field speed, flow, and related information.
• Reconstructed legacy code, notebooks, and reports into traceable ADC-to-g, FFT, SNR, cutting-point, and frequency-domain workflows, delivering executables/manuals while protecting company algorithm IP.
• Designed offline research topics, learning material, and starter notebooks, reviewing data assumptions/direction through weekly reports; the intern retained authorship of their analysis. Also owned company-side technical drafting, version convergence, and interface/checkpoint alignment for a smart-city edge-platform proposal, which remained an unsubmitted draft.

Data Science Intern, Institute of Sociology, Academia Sinica | Taipei, Taiwan | May 2019–Dec 2020

• Combined official sample addresses, point-to-point/fallback distances, survey-area grouping, OR-Tools TSP, and Folium for TSCS field routes, dispersion allowances, and budget estimates.
• Used Selenium, multiprocessing, Tor multi-port routing, and Folium for distance collection and GIS outputs as the project's contemporary technical approach.

MCIENet | M.S. Thesis, 2024

• Predicted 3D chromatin interactions/CTCF loops from DNA sequences using multi-scale CNN/Inception architecture, examining signals from different receptive fields.
• Independently led research decomposition, model and data-pipeline integration, experiment design/execution, ablations, cross-dataset evaluation, XAI, thesis writing, and reproducible repository engineering.
• Used BaseCNN comparisons, sequence-length settings, and ablation configurations to examine design choices. Retained DeepLIFT, ChINN/DNABERT2 comparison entry points, and shared split configurations for traceability. Findings are limited to completed experiments; comparison entry points and shared configurations were retained for further reproduction.
• Treated dataset, positional, and resolution differences as generalization checks, retaining cross-cell-line limitations. Adapted parts of positive/negative preprocessing from ChINN and distinguished that prior work from my model, experimental, and engineering contributions.

Financial-Text Sentiment Analysis | TCSE Research, 2020

• Studied stock-article sentiment using a Hierarchical Attention Network and sentiment lexicons. The work began with an approved undergraduate research grant and led to a TCSE publication.
• Performed all research work for the study; the paper was coauthored, so this scope does not imply sole authorship.
• Early repeated model changes yielded limited improvement. Gradually added data and parameter comparisons, retaining comparison conditions and experimental results as the basis for decisions.

Dynamic Embedding for Passenger-Flow Estimation | IIAI AAI Research, 2021

• Explored dynamic embedding for passenger-flow estimation. Responsible for method ideation, data processing, model construction, and experiment execution.
• Used data and parameter comparisons to examine model choices. The research received an IIAI AAI Competitive Paper Award; publication credit remains shared.

Research Experiment Practices Developed Over Time | 2020–2024

• Progressed from repeated model changes with limited improvement toward data and parameter comparisons, then more complete ablation design and parameter-adjustment logic in the MCIENet thesis.
• Used Shell to organize parallel experiments under limited GPU resources, including running two parameter configurations on one GPU with separate logs and checkpoints.
• Saved preprocessed data as H5, used early stopping, and first sampled three parameter values to inspect the curve before deciding how to adjust the search.
• Developed these practices across the 2020 sentiment study, 2021 passenger-flow study, PropGMAN, and MCIENet, with different levels of use in each project. This does not imply an identical complete pipeline across all projects or a measured overall speedup.

Dynamic Traffic Network Representation | KBS Collaborative Research, 2024

• Coauthored research on dynamic traffic network representation for passenger-flow prediction. My actual contribution was primarily supporting code and technical assistance.

Horus Human Identification System | Undergraduate Capstone, 2020

• The team integrated multi-object tracking, person re-identification, and face recognition into a multi-camera identification and tracking system.
• Primarily owned ReID feature-extraction integration, folder sampling/matching, cosine similarity, dual-threshold decisions, and tracking-record queue orchestration.
• Integrated the third-party FastReID base model with the team’s MOT and face-recognition modules. The team won first place in Soochow University's undergraduate capstone competition.

PropGMAN Asset Activation | Team Project, 2022

• In a three-person team, applied spatiotemporal modeling to property-transaction records to predict regional land prices and transaction volumes for asset-activation analysis.
• Proposed the project direction, led team coordination and division of work, and handled model implementation, experiments, and the final-round report. Retained preprocessing, model settings, experimental results, and reproduction instructions. The base GMAN method used an existing third-party research implementation. The team received a Creative Award in the AI for Asset Activation Innovation Contest.

Dynamic Bus Route and Schedule Planning | Project period to be confirmed

• Integrated tourism demand, Google Trends, Popular Times, attraction locations, and travel distances for route and schedule planning in eastern Taiwan. Built demand-allocation and distance-matrix workflows, explored OR-Tools TSP routes, and visualized routes and schedules with Folium, Gantt, and Sankey views.
• This was an exploratory independent project, separate from the Academia Sinica survey-routing work. The project period remains unconfirmed; no transit-operator adoption or operational outcome is claimed.