Bob Zheng

I work where large language models meet real software: agents that act through typed tools, research on aligning models at inference time, the ML infrastructure underneath, and the products that put it all in front of people.

Through roles at Carta, TD and Eon Media, and research with Prof. Edith Law (UW HCI Lab) and Prof. Usman Naseem (Macquarie University), I’ve shipped text-to-SQL data tooling, multi-agent prototypes, media-understanding pipelines and an accessibility scanner. I also first-authored VISPA, a paper accepted to the EMNLP 2026 Main Conference.

1. Software Engineer

   Carta

   ActiveData Ecosystem · Waterloo, ON

   Sep 2026 – Dec 2026

   • Improving data quality for a Snowflake Cortex text-to-SQL agent served over MCP: profiling client queries, enriching semantic views with metrics, dimensions and verified questions, and building benchmarks and regression tests.
   • Streamlined API-docs CI/CD with CircleCI validation, automated upstream PRs and Jenkins auto-deploy on merge.

   Snowflake Cortex · Text-to-SQL · MCP · CircleCI · Jenkins

2. Undergraduate Research Assistant

   UW HCI Lab

   ActiveWith Prof. Edith Law · Waterloo, ON

   May 2026 – Present

   • Investigating student-directed multi-agent orchestration for learning: shipped a platform where a student coordinates a persistent study companion and per-subject tutor agents against self-defined goals.
   • Designed the conversation-history layer that keeps long sessions in context: rolling compaction folds old turns into a running summary behind an advancing watermark, over a scoped memory store the agent mutates through remember / recall / update tools.
   • Made the LLM an actor over the domain via typed tool-calling (quest-board writes, memory, workspace files with PDF extraction, assessment logging), so agent turns persist real state rather than text.

   Multi-agent systems · Tool calling · Agent memory · HCI

3. Software Developer

   TD

   CompleteEnterprise Innovation · Toronto, ON

   Jan 2026 – Apr 2026

   • Built an Accessibility Scanner browser extension in TypeScript (Vite, Express.js) to automate enterprise-wide web accessibility and compliance checks across internal web properties.
   • Architected a PostgreSQL database with optimized indexing for scan-result retrieval and eligible-domain lookup; maintained the Jenkins CI/CD pipeline with Vitest unit tests, coverage thresholds and security scans.
   • Contributed to a proof-of-concept multi-agent design system on Google ADK and Anthropic models, where ADK agents coordinate via A2A and function calling to delegate sub-tasks to specialized agents.
   • Built collaboration features on GCS and Firestore and enabled bidirectional sync between code prototypes and Figma via MCP and REST, with custom parsers that compressed Figma context by 50% for LLM consumption.
   • Studied ROS and experimented with a humanoid service robot’s 2D LiDAR and sonar sensors, tuning costmaps and path planning to prototype in-person scenarios such as customer greeting and appointment check-in.

   TypeScript · PostgreSQL · Jenkins · Google ADK · A2A · MCP · Firestore · ROS

4. NLP Research Assistant

   Macquarie University

   CompleteWith Prof. Usman Naseem · Remote

   Sep 2025 – Aug 2026

   • First-authored VISPA, a training-free pluralistic alignment framework accepted to the EMNLP 2026 Main Conference, replacing costly multi-model ensembles with modular, value-steered generation on a single frozen backbone.
   • Engineered the inference system: DeBERTa-based NLI classifiers rank human values, then PyTorch latent-space manipulation injects value directions into intermediate LLM activations at inference time.

   PyTorch · Representation engineering · DeBERTa · NLI · Alignment

5. AI / ML Engineer, Co-op

   Eon Media

   CompleteToronto, ON

   Sep 2024 – May 2025

   • Optimized face and logo detection with YOLOv8 and Dlib, improving media-indexing accuracy and speed.
   • Shipped transcript chunking with all-MiniLM-L12 and fine-tuned Mistral-7B to summarize media into searchable segments.
   • Built a Flask backend for pipeline health monitoring and MLOps: dataset management and automated training and evaluation.
   • Cut AWS EKS pod-creation and runtime cost through container right-sizing, algorithmic optimizations and leaner AWS API usage; improved scalability with Docker, Kubernetes and ECR/EKS.
   • Developed multi-player tracking for sports video using Deep SORT with a custom ImageNet-based feature extractor, plus team and jersey-number recognition and key-moment detection.
   • Led a proof of concept for Cumulus Media: word-level-timestamped transcription with Whisper, radio diarization with pyannote, and speaker clustering and identification with scikit-learn to unlock voice-based asset monetization.

   YOLOv8 · Mistral-7B · Flask · Kubernetes · AWS EKS · Deep SORT · Whisper · pyannote

6. HCI Research Assistant

   University of Waterloo

   CompleteWaterloo, ON

   Sep 2024 – Dec 2024

   • Designed and piloted a user study with a Figma prototype to evaluate curiosity-driven learning and navigation behaviours.
   • Engineered PCuri, a Chrome extension (Node.js, WXT) that turns browsing history into a knowledge graph: KeyBERT keyword extraction, near-duplicate merging with SequenceMatcher, and edges from thresholded co-occurrence.
   • Integrated Vercel AI to generate research-question prompts and chat guidance that reference specific graph nodes.

   User studies · Figma · WXT · KeyBERT · Knowledge graphs

7. AI Software Developer Intern

   JuniorKids

   CompleteRemote

   May 2024 – Sep 2024

   • Gathered Reddit posts and comments via API and filtered them with spaCy named-entity recognition; fine-tuned a BERT model on historical posts for sentiment scoring.
   • Configured an LSTM to predict stock trends from sentiment scores, price trends and financial indicators.
   • Deployed scheduled cloud functions that store financial and social-media data in PostgreSQL and MongoDB.
   • Integrated RAG into a chatbot that answers in natural language using the latest market news and insider activity.
   • Earlier in the term: forecast clothing demand with LSTM models and streamlined competitor research across 500+ brands with the Similarweb API.

   BERT · spaCy · LSTM · RAG · PostgreSQL · MongoDB

For internships, please email my Waterloo address. For anything else, either address works.

• Primary emailb63zheng@uwaterloo.ca
• Personal emailshenyanzheng00@gmail.com
• GitHubgithub.com/BobZheng00
• LinkedInin/shenyan-zheng
• ResumePDF, October 2026