Applied AI Engineer

Walker Hyman

Forward Deployed Engineer at Ciridae.

Data Science & Philosophy at William & Mary.

Walker Hyman
Experience

Ciridae

Forward Deployed Engineer
Jun 2026 — Present
  • Embed directly with real-economy businesses to transform core operations into AI-native software, shipping production systems on multi-week timelines.
  • Build full-stack AI systems end to end — from requirements gathering with business stakeholders through production deployment.
  • Develop LLM orchestration and workflow automation that turn disconnected manual operations into working software.

Babel Street

Applied AI Engineering Intern
May — Aug 2025
  • Architected a production multi-agent deep research system—cut report generation from hours to under minutes.
  • Reduced irrelevant RAG results by 60% with a semantic-keyword hybrid retrieval pipeline.
  • Built tracing and observability infrastructure (prompt management, eval pipelines, user-level metrics) to accelerate iteration across agent workflows.

Centra Logistics

Software Engineer
Feb 2023 — Aug 2024
  • Built a novel unified AI commercial real estate property management and forecasting system.
  • Created CRE project management tooling to track costs, status, and progress across active projects.
  • Shipped forecasting tools for revenue projections and market analysis.
Skills
AI & ML
Multi-Agent SystemsNLPPyTorchHugging FaceOpenAI Agents SDKPydanticW&B
Full Stack
PythonTypeScriptReactDjangoFastAPIPostgreSQL
Infrastructure
DockerKubernetesTerraformAWSGCPRedisTemporalSLURM

Projects

03

Twitter Bot Detection via Graph Transformers

Multi-modal bot classifier on TwiBot-22 (1M users, 340M edges) fusing heterogeneous graph transformers, LoRA-fine-tuned XLM-RoBERTa text embeddings, and 23 engineered behavioral features. ~90% accuracy.

PyTorchPyGHuggingFacePEFT/LoRASLURM
04

VAE Face Generation on CelebA

Standard VAE and VQ-VAE for face image generation and reconstruction on 200K+ CelebA images. VQ-VAE uses a learned codebook with convolutional encoder/decoder. Evaluated with FID, PSNR, and MSE.

PyTorchTorchvisionpytorch-fidSLURM
05

LSTM Code Summarization

End-to-end pipeline for Java method summarization: mines ~50K method/Javadoc pairs from GitHub, tokenizes with CodeT5+, and trains a seq2seq LSTM with attention. Evaluated on BLEU-1.

PyTorchHuggingFaceCodeT5+GitHub API
06

Qwen RAG for Java Bug Fixing

End-to-end pipeline: trains a SentencePiece tokenizer on CodeSearchNet, pre-trains T5-small with span corruption, fine-tunes on bug fixing, and benchmarks against Qwen2.5-Coder zero-shot and CodeBERT-retrieval RAG.

PyTorchHuggingFaceT5QwenCodeBERTSentencePieceSLURM