Work

Outcomes, Not Slide Decks

Below are indicative engagements—names and sectors anonymised—spanning customer-facing AI experiences and MCP-backed access to heritage systems. Every build pairs UX clarity with integration discipline so wins survive contact with real operations.

MCP & legacy focus → Mobile app delivery →

B2B SaaS · Customer support

Support Copilot Rollout

Challenge

Support volume was climbing but headcount was flat. Agents were copying answers from three internal wikis and a ticketing system with inconsistent tagging—quality varied wildly shift to shift.

Approach

We built a retrieval pipeline over approved macros and policy docs, routed high-risk categories to human review, and integrated ticket creation through governed API calls—not raw database access.

Outcome

Reduced average response time by 38% while increasing answer quality consistency across shifts.

Stack: LLM routing, retrieval, policy filters, human handoff

Enterprise sales · CRM-heavy

Sales Intelligence Assistant

Challenge

Reps spent hours before every call stitching account history from CRM notes, email threads, and a legacy pricing spreadsheet that only one person understood.

Approach

MCP tools wrapped CRM read/update operations with role-based scopes. A morning briefing agent summarised open opportunities and flagged stale records for cleanup.

Outcome

Automated account prep and CRM updates, saving 11+ hours per rep each week.

Stack: MCP connectors, workflow agents, CRM actions, observability

Distribution · Field operations

Legacy ERP MCP Gateway

Challenge

Field teams needed inventory and fulfilment status on site visits, but the 15-year-old ERP only exposed brittle SOAP endpoints—and IT refused to hand SQL credentials to any AI experiment.

Approach

We implemented an MCP tool layer with row-level auth mapped to territory, structured logging for every lookup, and explicit rate limits. No model saw raw SQL.

Outcome

Field teams query inventory, fulfilment status, and pricing through an assistant—without bypassing business rules baked into the ERP.

Stack: MCP tool layer, SOAP/REST adapters, row-level auth, structured logging

Operations · Regulated industry

Internal Knowledge Agent

Challenge

SOPs lived in SharePoint, Confluence, and PDF binders on a shared drive. New hires asked the same questions in Slack channels that senior staff had answered dozens of times.

Approach

Indexed approved sources with citation UX, freshness checks on policy documents, and an eval harness tuned to “must not hallucinate compliance steps.”

Outcome

Gave operations teams a single AI interface to SOPs, docs, and live business data—with source links on every answer.

Stack: RAG pipeline, vector search, source citation, eval harness

Want a similar outcome?

Share the workflow that burns time today and the systems we would need to respect. We will tell you plainly if MCP, RAG, or a simpler integration is the right first slice.

Start a conversation