Shawn YapEmail

Shawn Yap Zheng Yi

Solution Architect, AI Systems · BeyondEdge · Singapore

I design and ship production AI systems: agent platforms, grounded retrieval, and speech pipelines, each one gated, validated, and measured against a baseline.

  1. User message

  2. Scope gate

    Off-scope queries stop here

  3. Query rewrite

    Memory-aware, conversational state

  4. Orchestrator

    9 intents

  5. Tiered retrieval

    4 Qdrant collections

    • Exact, then partial, then semantic
    • Geocoded proximity filters
    • User-confirmed search slots
    • API-sourced numeric facts

  6. Qwen3.5-9B + LoRA

    Tuned on ~1k client-curated conversations

    Reply fidelity +13% over base model

  7. Grounded recommendation
View A · Property AI AssistantConsumer assistant for a market-leading property agencySchematic. Components from production build; stage order simplified.

Results, drawn to scale

Numbers from shipped systems and finished research. Where the resume gives absolute values they are plotted on a ruler; relative gains are listed as notes.

Detail B1 · Whisper + LoRA word error rate

010203040−35 ptsBefore 40%After LoRA 5%
FCS Solutions, voice pipeline

Detail B2 · BERT multi-label POFMA actor classification

0.700.750.800.850.90F1 0.781Accuracy 0.847Scale 0.70–0.90
Final-year project, NTU

Detail B3 · LLM risk reviewer, trades rejected

Unbounded · 32 of 32 rejectedCapped, risk-raise only · 11 rejectedHatched = rejected trade
Survivor trading system

General notes

ValueResultWhere
1.30%EER improvement in audio deepfake detectionHTX
2.+13%Reply fidelity over base after LoRA fine-tuneBeyondEdge
3.200+Intent classes with calibrated confidence thresholdsFCS Solutions
4.676Backtest sessions verifying no look-ahead biasSurvivor
5.5k+Point-in-time forecasts scored with Brier calibrationSurvivor
6.560Page LLM-wiki behind a WhatsApp second-brain agentBeyondEdge

Experience

Revision history, newest first. Client work is described, not named.

  1. BeyondEdge

    Solution Architect, AI Systems (current role)

    Jul 2026 – PresentSingapore

    Property AI Assistant

    Client withheld · Market-leading property agency

    • Led the build of the consumer-facing assistant, architecting its orchestration pipeline across 9 intents and co-developing backend API endpoints, with scope gating, conversational state management, and memory-aware query rewriting.
    • Engineered tiered retrieval across 4 Qdrant collections, falling back to partial to semantic matches, with geocoded proximity filters, user-confirmed search slots, and API-sourced numeric facts for grounded property recommendations.
    • Fine-tuned its Qwen3.5-9B model with LoRA on ~1k client-curated conversations, lifting reply fidelity 13% over base.

    Enterprise Agentic AI Platform

    Client withheld · Arm of a major global investment group

    In production

    • Led the build and production deployment of a multi-tenant LLM agent platform with LiteLLM model routing, per-team token budgets, policy-gated MCP tools, and data-classification controls on GKE.
    • Prototyped a WhatsApp second-brain agent that distils free-text feedback into persistent memory over a 560-page LLM-wiki, productionising it into configurable department-level agents non-technical staff can tune without code.
    • Grounded agent briefs with per-claim citations to archived sources, prompt guards against forged citations, and watchdog alerts replacing silent run failures.

  2. FCS Solutions

    AI Engineer

    Jul 2025 – Jul 2026Singapore

    • Productionised an ASR–LLM–TTS pipeline with tool routing, schema validation, confidence gating, and telemetry.
    • Fine-tuned Whisper with LoRA, reducing WER from 40% to 5% and improving domain-specific intent recognition.
    • Built and fine-tuned a 200+ class intent-classification system with calibrated confidence thresholds for production routing.
    • Built schema-aware NL-to-SQL agents with constrained decoding, query validation, retries, and unsafe-query blocking.

  3. Home Team Science and Technology Agency (HTX)

    Data Scientist & Machine Learning Engineer Intern

    Oct 2024 – Jun 2025Singapore

    • Led R&D on audio deepfake detection using speech embeddings and prosodic features, improving EER by 30%.
    • Built reproducible, seed-fixed benchmarks, evaluating via EER, ROC-AUC, precision-recall, and failure-mode analysis.
    • Designed denoising and augmentation pipelines, sweeping EER across SNR levels to characterise low-SNR robustness.
    • Prototyped LangChain LLM and RAG workflows for summarisation and contextual retrieval over internal documents.

Projects

Survivor

Evolutionary multi-agent trading system

Jun 2026 – Present · Python · PostgreSQL · Docker · LLMs

  • Breeds, tests, and retires trading strategies by genetic search over 16 families and 8 regimes.
  • Scored 5k+ point-in-time forecasts against realised outcomes with Brier calibration, demoting overconfident strategies.
  • Capped the LLM reviewer to raising risk only; unbounded, it rejected all 32 trades (vs. 11) and halted trading.
  • Runs self-hosted Qwen inference under schema-constrained decoding with Pydantic validation of every model output.
  • Eliminated look-ahead bias structurally so backtests cannot access future data, verified over 676 backtest sessions.

POFMA actor classification

Final-year project: transformer classification and sentiment analysis

Jan 2024 – Dec 2024 · PyTorch · BERT

  • Fine-tuned BERT for multi-label POFMA actor classification over 512-token chunks, at 0.847 accuracy and 0.781 F1.
  • Built scraping and PDF pipelines with contextual feature engineering, plus LLM sentiment analysis of public reaction.
Source on GitHub

Referenced repositories

RepositoryDescriptionLang
quantwachstumSelf-hosted equity research triage: transparent scoring, scenario ranges, Telegram delivery.Python
pofma-sentiment-analysisActor classification and sentiment analysis over every POFMA notice since 2019.Python
htx-dl-denoiseNoise reduction with DCCRNet and traditional audio engineering.Python

Skills

Parts list. Grouped by where each is used, not by how recently it was learned.

ItemGroupParts
1Machine Learning
  • PyTorch
  • Hugging Face Transformers
  • TensorFlow
  • LoRA/PEFT
  • Fine-tuning
  • BERT
  • Qwen
  • Whisper
  • NLP
  • ASR
  • Embeddings
  • Dataset Curation
2Agentic AI & LLM Systems
  • Tool Calling
  • Tool Routing
  • MCP
  • RAG
  • Hybrid Retrieval
  • Qdrant
  • Structured Outputs
  • Constrained Decoding
  • NL-to-SQL
  • Prompt Engineering
  • LangChain
  • OpenClaw
3LLM Reliability & Evaluation
  • Guardrails
  • Confidence Gating
  • Schema/Execution Validation
  • Retry Logic
  • Calibration
  • Threshold Tuning
  • WER
  • EER
  • F1
  • ROC-AUC
  • Brier Score
  • Error Analysis
  • Benchmarking
4Data Science
  • pandas
  • NumPy
  • SQL
  • Feature Engineering
  • Statistical Analysis
  • Experiment Design
  • Backtesting
  • Point-in-Time Evaluation
5AI Engineering
  • Python
  • Async Python
  • FastAPI
  • REST APIs
  • SQLAlchemy
  • PostgreSQL
  • Docker
  • Git
  • AWS
  • GCP
  • Azure
  • Pytest
  • Automated Testing
  • Model Deployment
  • Claude Code

Education & languages

Education

  1. Nanyang Technological University

    Bachelor of Computing (Hons), Computer Science, Minor in Modern Languages

    Singapore · Aug 2021 – May 2025

  2. University of Stuttgart

    Semester exchange and German immersion programme, Computer Science

    Germany · Mar 2024 – Aug 2024

Languages

English
Native
Mandarin Chinese
Native
Cantonese
Working proficiency
German
CEFR B1
Spanish, Italian
Learning

Certifications

  1. DeepLearning.AI TensorFlow Developer Specialization

  2. Google Advanced Data Analytics Certificate

Contact

shawnyzy1@gmail.com
Email ShawnLinkedInGitHub
Drawn by Shawn YapSingaporeSheet 1 of 1 · Rev C© 2026