I embed with business teams and take agentic AI from discovery to a production system they own. Currently acting lead Forward Deployed Engineer in TELUS Digital's AI Accelerator FDE program.
Sanitized versions of production systems from client and internal engagements. All client identifiers, data and credentials are removed.
| Project | What it is | Stack |
|---|---|---|
| website-visitor-lead-check-agent | Current FDE engagement. An agent that turns ~3,000 identified website visitors a week into a short, explainable shortlist across 127+ key accounts: rules decide and write each reason and next step, Gemini explains. Includes per-user isolation from verified IAP tokens, memory-only personal data, 13 business-editable rule lists, and hub-and-spoke target architecture diagrams. | Google ADK · Gemini (Vertex AI) · Cloud Run · IAP · pandas · pytest |
| ai-assistant-langgraph-backend | Conversational AI backend for a major media publisher: a per-request LangGraph agent with LLM-as-judge guardrails, intent routing, grounded RAG answers with article recommendations, multi-brand personas and a GDPR deletion worker. 304 tests, 95% coverage. | LangGraph · FastAPI · Claude / Gemini · DynamoDB · Redis · Kubernetes |
| multi-agent-support-platform | A Slack-native multi-agent support platform for industrial pipeline operations, plus a Document AI pipeline that fills engineering workbooks from RFI documents, evaluated against ground truth. | Google ADK · Gemini · Cloud Run · BigQuery vector RAG · Document AI |
| email-qa-llm-pipeline | An LLM email-QA system that scans 40k+ marketing emails for 15 issue types, with parallel Bedrock workers, checkpoint/resume, versioned prompts and a precision/recall evaluation harness. | AWS Bedrock (Claude) · S3 · pandas · boto3 |
| llm-sentence-generation-eval | A dual-LLM benchmarking platform (GPT-4.1 vs Gemini 2.5) with parallel generation over CEFR-levelled data, token and cost accounting including thinking tokens, and an 8-metric NLP evaluation harness. | OpenAI · Vertex AI / Gemini · scikit-learn · NLTK |
More:
- AI_news: a personal daily AI brief that collects, ranks and summarises news with Gemini.
- blu_baar_chatbot: a bilingual chatbot API on Cloud Run with Cloud Build CI/CD.
- aml-assistant: a compliance RAG chatbot.
- CBC-news-RAG: RAG over news articles and editorial guidelines.
- receipt_scanner: a serverless Document AI workflow on GCP.
- respondelligent_cpu: mBART fine-tuning and deployment.
- code_portfolio_examples: classic ML and SQL samples.
- Agent frameworks: Google ADK (agents, tools, plugins, A2A) · LangGraph · LangChain
- Models: Gemini (Vertex AI) · Claude (AWS Bedrock) · OpenAI GPT · LiteLLM gateway
- Patterns: RAG · guardrails · intent routing · rules + LLM hybrids · prompt engineering
- Evaluation: ground-truth test suites · precision / recall harnesses · LLM-as-judge
- Google: Gemini Enterprise · Vertex AI Agent Engine
- Salesforce: Agentforce (Agent Builder, Prompt Builder, Flow / Apex actions) · Data 360 · Einstein Trust Layer · Testing Center
- GCP: Cloud Run · GKE · Vertex AI · BigQuery · Document AI · Cloud Build · Artifact Registry
- AWS: Bedrock · S3 · DynamoDB · SageMaker
- Infrastructure: Terraform · CI/CD · Workload Identity Federation · hub-and-spoke projects
- Access: Identity-Aware Proxy (signed JWTs) · IAM · least-privilege service accounts
- Data: privacy by design · memory-only personal data · data-enablement reviews
- LLM safety: prompt-injection hardening · content security policy
- Languages and APIs: Python · FastAPI · SQL · JavaScript
- Data: pandas · scikit-learn · Redis / Valkey · DynamoDB · BigQuery
- Ops: Docker · Kubernetes · pytest · Datadog · Sentry · OpenTelemetry
- Discovery: workshops · use-case prioritisation · data-access mapping
- Design: solution proposals · architecture decision records · demos for leadership
- Execution: stakeholder management · Jira · Gantt planning · weekly status reporting
LinkedIn · ovsyannikovilyavl@gmail.com
MSc; two IEEE publications on deep learning for computer vision.