I design and build systems that turn manual processes into reliable operations.

Problem

The business was missing a significant number of enquiries outside operating hours, resulting in lost leads and revenue the team had no visibility over.

Decisions

• Prioritise full coverage across both phone and email channels
• Ensure no enquiry is lost, delayed, or ignored
• Maintain communication quality consistent with a human response
• Design for real-time response with structured follow-up

Solution

• Built a multi-channel system integrating a voice agent (Vapi), email automation, and enquiry routing logic
• Implemented structured data parsing to extract lead details from both channels
• Introduced fallback handling for edge cases and unresolved queries
• All enquiries feed into a tracked pipeline with logging and status visibility

Controls

• Output parsing before any routing decisions are made
• Deterministic logic controlling response paths
• Error logging and notification workflows
• Human fallback triggered for enquiries that cannot be resolved automatically

Outcome

• $60,000 in additional booked jobs in the first month
• 100% after-hours enquiry coverage across voice and email
• Zero missed inbound opportunities during the coverage window

Problem

High volume of structured email communication required manual handling, consuming significant staff time and creating delays across the team.

Decisions

• Avoid broad tool integrations to reduce hallucination risk
• Build a lean, controlled API layer rather than relying on general LLM knowledge
• Prioritise reliability and consistency over flexibility

Solution

• Built a custom email agent system with a tailored API tool layer (POST/GET endpoints)
• Integrated Pinecone for structured semantic retrieval against case documents
• Designed system to handle routine check-ins, coordination, and escalations
• Human-in-the-loop approval step before any email is sent

Controls

• Structured output validation before responses are generated
• Confidence-based routing for uncertain cases
• Human review triggered for low-confidence or sensitive responses
• Retrieval validation before any response is drafted

Outcome

• Reduced manual communication workload by 20+ hours per week across 3 staff
• Improved response consistency across all client communications
• Hallucination rate reduced by approximately 20%
• Retrieval accuracy improved by approximately 25%

Problem

Broker submissions required manual review and transformation into structured NBI documents, causing delays and creating audit gaps.

Decisions

• Maintain strict validation before any processing occurs
• Ensure the system communicates errors clearly to users
• Track all actions for auditability from the start

Solution

• Built an email-triggered automation system that intercepts and processes broker submissions
• Extracted and structured document data using LLM with defined output schemas
• Generated NBI outputs automatically against required templates
• Implemented SQL-based logging for full transaction traceability

Controls

• Multi-step validation gates before AI processing begins
• File type and size validation at ingestion
• Structured error responses sent back to users for invalid submissions
• Full logging of every system action for audit purposes

Outcome

• 80% reduction in processing time per submission
• 90% of the workflow automated end-to-end
• Increased reliability and consistency across all submissions
• Clear, queryable audit trail for all operations

Problem

The existing system was poorly scoped, with inconsistent data structures and requirements that continued to change throughout the engagement.

Decisions

• Redesign the system rather than patch the existing one
• Clean and restructure the data before applying any automation
• Adapt systematically to ongoing requirement changes without losing progress

Solution

• Rebuilt the database structure to support reliable automation
• Implemented OCR-based document processing for purchase order ingestion
• Introduced vectorisation for structured data handling
• Created a booking automation system on top of the new data foundation

Controls

• Data validation at ingestion before processing
• Structured data pipelines with clear state tracking
• Controlled output generation with defined schemas
• Incremental rollout to reduce risk across phases

Outcome

• Delivered the full system despite sustained scope change over 6 months
• Significantly improved data quality and usability
• Established a reliable, scalable foundation for future automation

Problem

High volumes of routine customer enquiries over the phone were creating wait times, inconsistent responses, and operational load on staff.

Decisions

• Design for real-time voice interaction, not delayed or async responses
• Integrate directly with internal APIs for accurate, live data
• Introduce verification before sharing sensitive information
• Ensure clean escalation to human agents when needed

Solution

• Built a modular voice agent capable of understanding natural language queries and retrieving structured data via API
• Integrated Vapi for call handling, ElevenLabs for voice synthesis, and an LLM for intent and response generation
• Implemented an MCP server to enable real-time API interaction during calls
• Designed full conversation flow covering order tracking, availability, and store information

Controls

• Real-time verification before exposing any sensitive order or account data
• Fallback logic for queries the agent cannot resolve
• Escalation to human agents for complex or sensitive cases
• Full interaction logging and metadata tracking for every call

Outcome

• Reduced routine customer support load on staff
• Improved response speed and consistency across enquiry types
• Enabled scalable voice-based customer service without additional headcount
• Human fallback maintained for edge cases

Problem

Claims documentation required manual transcription, data extraction, document creation, and CRM updates — creating delays, inconsistency, and significant manual workload.

Decisions

• Use the meeting transcript as the single source of truth for all downstream outputs
• Extract structured data via LLM with defined output schemas
• Ensure all outputs align with compliance templates before any CRM update
• Integrate directly into the CRM to eliminate re-entry entirely

Solution

• Built a workflow that captures Google Meet transcripts, extracts structured claim data using an LLM, and generates completed claim documents
• Integrated directly with Zoho CRM to update records automatically from extracted data
• Used Gemini for structured extraction with schema-based output validation
• Documents generated against standardised compliance templates

Controls

• Schema-based extraction with structured output validation
• Low-confidence cases routed to human review before CRM update
• Standardised document templates to ensure compliance consistency
• Full data traceability from transcript through to document and CRM record

Outcome

• Eliminated manual data entry from the claims documentation process
• Increased speed of claim processing end-to-end
• Improved documentation consistency and compliance
• Significantly reduced operational workload for the claims team