I build agents that do the work.

An agent factory for people who don't code

I built an Agent Factory on GitHub Copilot with custom MCP servers for the bank’s on-prem Jira, Confluence, Bitbucket, Jenkins and Edge Node, so non-technical staff can build their own agents. Two that came out of it: a scrum-master agent that folded three sprint meetings into one, and a Confluence knowledge agent that searches the documentation and writes updates back.

Migrating 14,000 views with agents

Moving a bank’s consumption layer from Synapse to Snowflake: ~14,000 views, ~2,000 pipelines and ~500 repositories. Agentic CI/CD workflows promoted the repos from Dev to SIT to Prod and cut pipeline runs from ~2 hours to ~10 minutes. A validation utility, from metadata checks to sorted-hash record comparison, surfaced a defect pattern across ~18% of the estate, and an agent-driven fallback during a Control-M outage saved about three weeks.

Debt collection, governed by design

I led the build of a multi-agent debt collection system for a bank: an orchestrator, a policy hub and an audit platform on GCP. What makes it different is the governance around the conversation, not just the conversation itself.

• Policy as a constraintThe policy layer works out each customer's eligible hardship offers before the agent writes a word, so it can only offer what policy allows.
• AuditableEvery turn is recorded with its intent, sentiment, reasoning and the policy that governed it.
• ObservableEach conversation can be traced end to end across the orchestrator, agents and policy hub.
• Conversation and analysisAgents hold the conversation and analyse it as it happens, from intent to sentiment.

From regex to hybrid RAG

I moved MogoPlus’s flagship transaction categorisation engine from regex matching to a hybrid RAG system with sub-agents. Merchants are resolved against the government business registry, a deep-search agent writes a description of each business, and pgvector retrieval with an LLM reranker maps it to an industry code. It recovered about 80% of the transactions that were stuck in the “other” bucket.

Documents in minutes, not hours

OCR-based document processing wired into the agent workflows, so documents are read, their fields extracted and handed to the agents without manual handling. Processing time dropped from about an hour to about five minutes per document.