← Selected Work

Kaddi

AI sales execution platform · Technical co-founder & sole engineer

A $2M-funded, AI-native sales execution platform that turns live deal data — emails, call transcripts, documents and CRM records — into deal guidance and strategic pipeline intelligence. Built from an empty repository and operated in production for five years, owning architecture, code, security and operations end-to-end.

kaddi.io →

5 years

In production, from empty repository to live sales deals

84%

Reduction in LLM cost through token budgeting and model selection

11 tools

Available to autonomous agents across up to 10 turns

2023

Australian iAwards finalist

Highlights

Agent-based workflows. Autonomous multi-turn tool-calling agents with up to 11 tools and 10 turns, alongside deterministic orchestration for repeatable pipelines.
Production LLM features used in live deals — deal guidance, RFP automation and buyer messaging, with streaming responses, structured outputs, Whisper speech-to-text and ElevenLabs voice synthesis.
Analysis beyond the context window. Thousands of emails, transcripts, documents and CRM records processed with asynchronous pre-compression and parallel map-reduce, grounding insight in real communications rather than hand-maintained CRM fields.
84% reduction in LLM cost through token budgeting, task-specific model selection and per-feature usage reporting.
Turned ambiguous customer and commercial problems into shipped AI features, often in days.
2023 Australian iAwards finalist.

Technical detail

ReactTypeScriptPythonDjango / DRFPostgreSQL + RLSAWSOpenAIRAGWhisperElevenLabsSalesforce

RAG pipeline over customer sales content: ingestion, chunking, embeddings, retrieval, reranking and prompt assembly, so answers are grounded in the customer's own material.
Multi-tenant by construction. PostgreSQL row-level security enforced at the database, with the tenant context carried explicitly into background tasks and threaded work that have no request to inherit it from.
Service-oriented Django backend. DRF views delegate to services; every external system — OpenAI, Salesforce, Google, S3 — sits behind its own boundary, and prompt construction is kept strictly separate from orchestration.
OAuth integrations with Google Sign-In, Gmail, Google Drive, Zoom and Salesforce, making deal data available to LLM workflows.
Asynchronous processing with SSE for ingestion, syncing and long-running AI tasks, so nothing blocks the interface.
Structured regression evaluations for model and prompt changes, run against the live LLM with results tracked over time.
Owned CI/CD, QA, releases, monitoring, security and AWS operations, including a parallelised test suite gating every merge.

Kaddi

Highlights

Technical detail

ADC Ready

Borderline