Technical Evidence Index
This index is the fastest route through the portfolio for a recruiter, sales leader, or Solutions Engineering reviewer. It separates direct professional delivery, collaborative platform outcomes, and independent portfolio builds so that every claim has an explicit boundary.
Public professional artifacts are sanitized reconstructions. Production code, customer or company data, credentials, internal identifiers, endpoints, prompts, and screenshots are excluded.
Evidence map
| Evidence class | What it proves | Best starting point |
|---|---|---|
| Direct professional delivery | Ability to build, productionize, operate, control, and explain enterprise cloud applications | GCP intelligence automation |
| Collaborative platform contribution | Ability to work across data pipelines, feature serving, model lifecycle, GPU serving, validation, and infrastructure economics | Enterprise MLOps and GPU serving |
| Direct governed AI application | Ability to constrain model use with approved data, deterministic calculations, query controls, and reviewable outputs | Governed FP&A agent blueprint |
| Independent technical-commercial builds | Ability to translate discovery evidence into sizing, architecture, validation, and financial decision support | Live Northstar walkthrough |
| Implementation quality | Ability to build runnable software with validation, persistence, exports, tests, CI, containers, and documented limitations | Independent workflow repositories |
Professional evidence
Production intelligence automation on GCP - direct delivery and operation
Contribution: built and operated seven intelligence workflows and two recurring reporting paths using Cloud Run, Cloud Scheduler, Cloud Storage, BigQuery, Vertex AI, Secret Manager, and SendGrid.
Controls: deterministic KPI calculation, readiness gates, resilient source handling, governed recipients and preferences, accepted-delivery checks, sent-history protection, and reversible promotion procedures.
Measured outcome: 60% faster reporting turnaround and 100+ annual hours saved within a wider reporting environment covering 45+ datasets and 80+ properties.
Enterprise MLOps and GPU serving - collaborative platform contribution
Contribution: Kafka and Airflow pipeline components; Feast, Redis, and Snowflake feature paths; shared NVIDIA Triton serving on AWS EKS; GPU optimization; performance validation; and technical-commercial analysis.
Global portfolio outcomes:
| Technical measure | Reported outcome | Decision relevance |
|---|---|---|
| Feature freshness | 24 hours to approximately four minutes | Supports more responsive operational and pricing decisions |
| Forecast MAPE | 12.5% to 4.5% | Establishes the forecasting improvement behind downstream pricing decisions |
| RevPAR and GOP | +4.2% RevPAR; +S$4.8M annualized GOP | Connects model and pricing performance to commercial results |
| Utility efficiency | 14% reduction; approximately S$1.8M OpEx savings | Connects LSTM meter-anomaly detection to operating cost |
| Asset reliability | 42% less catastrophic downtime; approximately S$1.5M deferred CapEx | Connects FFT/Weibull early-warning analysis to capital preservation |
| Maintenance service recovery | 77% lower MTTR; +3% repeat bookings | Connects BERT-assisted work-order routing to faster resolution and customer retention |
| GPU utilization | Approximately 5% to above 80% | Establishes productive infrastructure density |
| Serving scale | 10x demand below 150 ms p99 | Tests scale while protecting the service-level target |
| Hosting cost | 58% reduction; approximately US$240K annually | Connects serving architecture to run-rate economics |
These are collaborative global portfolio outcomes. They are not represented as the result of one contributor or one model acting independently.
Governed FP&A analytics agent - direct application build
Contribution: built a Google ADK and BigQuery workflow that turns approved finance and property-performance questions into source-labelled analysis.
Controls: approved objects, source-specific property resolution, read-only SQL, byte and result caps, KPI-semantic rules, and data-quality checks. The model plans and explains; BigQuery performs the calculations.
Measured outcome: common ad hoc extraction fell from hours to under two minutes.
Applied ML evidence
| Operating problem | Technical pattern | Published evidence level |
|---|---|---|
| Demand forecasting and bounded pricing | Temporal Fusion Transformer plus Deep Q-Network | Quantified global portfolio outcome |
| Utility anomalies and leaks | PyTorch LSTM autoencoder over 15-minute sequences | Quantified global portfolio outcome |
| Mechanical failure risk | Edge FFT features plus Weibull survival analysis | Quantified global portfolio outcome |
| Maintenance-ticket classification and dispatch | Fine-tuned BERT classification | Quantified global portfolio outcome: 77% lower MTTR and +3% repeat bookings |
Independent customer-decision workflow
The four workbenches implement distinct stages of one reviewable AI-infrastructure decision:
- Opportunity and Discovery separates sourced evidence from interpretation and recommends advance, reshape, nurture, or disqualify.
- Capacity and Commercial Sizing produces indicative compute, memory, storage, network, power, and commercial ranges with sensitivities and validation requirements.
- Solution Configurator produces an explainable architecture hypothesis with alternatives, risks, and approval gates.
- TCO and ROI compares operating models through deterministic calculations, evidence confidence, sensitivity, lineage, and review-ready exports.
The Northstar AE-to-SE handoff and live walkthrough show how the four stages connect without presenting fictional values as customer results.
Public-artifact standard
- Professional outcomes are labelled as direct workflow outcomes or collaborative global portfolio outcomes.
- Fictional scenario inputs remain marked illustrative in every output.
- Deterministic code owns scoring, sizing, calculations, state, and delivery decisions.
- Model-generated interpretation never silently changes a calculation or approval gate.
- Unknown evidence remains visible instead of being replaced by a precise-looking estimate.
- Architecture, benchmark, security, pricing, and finance approval remain explicit human review gates.