Organizational AI Enablement & Adoption Case Study
Ava: Enterprise AI Platform
The Head of Technology who led inception, fund-raising, team building, design, implementation, delivery and adoption of Ava, a governed enterprise AI assistant now cutting document turnaround time by up to 95% across a growing set of business functions.
Led a 5-person cross-functional team to build and ship Ava, an AI enterprise assistant that cut compliance response drafting time by 95% (3 weeks → ~1 day) and PDF-to-Excel clause conversion time by 79% (2 weeks → 3 days).
Architected a governed retrieval-augmented generation (RAG) pipeline on AWS (Cognito, S3, Bedrock, DynamoDB) processing 100+ enterprise documents to date, with zero customer data leaving controlled infrastructure.
Designed a human-in-the-loop publishing model where AI-generated drafts require explicit reviewer approval before becoming validated knowledge assets, eliminating ungoverned AI output as a risk in a compliance-sensitive workflow.
Drove adoption of Ava from a single-team Presales tool to a platform expanding into Engineering and Finance, with a technical architecture designed for reuse across business functions from day one.
Every product and architectural decision is deeply rooted in organizational culture and process
❓Why not just use ChatGPT / Copilot / an off-the-shelf tool for this?
Off-the-shelf tools can't ground responses in r2p's own proven prior work, can't guarantee proprietary tender content stays inside our infrastructure, and don't give us a publish boundary we control. Ava's value isn't "AI writes text". It's the retrieval layer and the governance layer (human approval, full audit trail) built specifically around a compliance-sensitive workflow.
❓How do you know it's not hallucinating?
Ava doesn't generate from nothing. Every draft is grounded in retrieved, previously approved responses, with citations attached showing exactly what it was based on. And nothing published is auto trusted: a human reviewer sees the draft and its sources before approving. Hallucination risk is mitigated architecturally (retrieval-grounded generation) and procedurally (mandatory human review), not just by hoping the model behaves.
❓What happens if a reviewer just rubber-stamps every draft?
This is addressed via process and governance such as spot-checking, sampling, or accountability tied to the reviewer's identity in the audit log.
❓Where do customer or proprietary data actually go?
Nowhere outside business's own AWS account. Embeddings are generated by a self-hosted Ollama instance on ECS, not a third-party embedding API and storage (S3), vector search (Qdrant), and state (DynamoDB) all live inside r2p-controlled infrastructure. No customer document content is used to train any third-party model.
❓What's the actual ROI or business case?
Compliance response drafting time down 95% (roughly 3 weeks to about 1 day), PDF-to-Excel clause conversion down 79% (2 weeks to 3 days) that results in cost savings and capacity freed up.
❓Why self-host embeddings instead of using a managed API; wasn't that more work?
Yes, it was more upfront engineering effort. But sending clause-level tender content to a third-party embedding endpoint is a data-residency decision made by default, whether intended or not. Self-hosting on our own ECS infrastructure keeps that decision explicitly in our hands, which mattered given the compliance-sensitive nature of the content.
❓What's next for Ava?
Engineering (user story generation, test case generation, knowledge base lookup) and Finance (forecasting, document workflows) are both in early stage, reusing the same governed retrieval architecture that proved out in Presales.