Four series of production-grade architecture. Each project addresses a real enterprise problem with a documented design, a named stack, and a traceable decision trail. No demos. No toys.
"The function of the Architect is well understood."
Each entry in this series applies the same architectural philosophy — TOGAF ADM, multi-agent orchestration, XAI-first ML, and HITL governance — to a different enterprise domain. Same rigour. Different problem. One coherent vision.
A complete, six-phase enterprise AI architecture for ClaraVis Medical Systems — a €1.2B medical imaging OEM operating in 34 countries. Eight intelligent modules. Every decision explainable. Every agent action auditable. EU AI Act, ISO 13485, ASC 606, and GDPR compliant — by structural design, not post-hoc reporting.
An explainability-first, human-gated documentation pipeline for technical writers. Five agentic stages, one human gate, zero black-box decisions. Google Style Guide-compliant first drafts from any intake format.
HR automation for the 2.7B deskless workers with no HR system. WhatsApp-native, voice-first, policy-governed RAG. Full leave management at $0.13 per employee per month. Works on any Android. Zero new apps.
Autonomous financial close, accounts payable/receivable, and regulatory reporting for CFO-level governance. ASC 606 / IFRS 15 compliant by write-path constraint. Every journal entry explainable, every recognition decision auditable.
Demand forecasting, supplier risk scoring, inventory optimisation, and logistics event processing on GCP. Real-time ML over streaming telemetry. Vertex AI pipelines with automated drift detection and retraining triggers.
Regulatory change monitoring, policy gap analysis, audit trail generation, and risk register automation for financial services and pharma. Compliance as a continuously evaluated property — not a quarterly report.
Three systems that don't just generate outputs — they operate across enterprise tools, make decisions within defined boundaries, and route to humans at every threshold that matters. Capture, knowledge, action.
A Contact Centre AI sales agent that handles the first 11 turns of an enterprise sales inquiry fully autonomously — qualification, product fit, configuration options, pricing estimate — before escalating to a Senior Account Executive with a complete briefing document the agent prepared. Turn 12 is always human.
Inbound enterprise sales qualification burns senior AE time on deals that aren't ready
CCAI + ADK agent · 11-turn autonomy boundary · human handoff at commercial terms
Gemini · Google CCAI · ADK · Dialogflow · Vertex AI
A swarm of specialist agents that decompose a research brief, execute parallel searches, synthesise findings across sources, resolve conflicts, and produce a structured enterprise-grade report — with full source attribution and confidence scoring throughout.
Receives a sales opportunity, autonomously pulls CRM data, checks pricing rules, retrieves contract precedents, calculates margin, generates a compliant quote, and routes for approval above threshold — posting back to Salesforce without human intervention below the defined deal value.
Three RAG systems that solve three different retrieval problems: security and data custody, precision and query optimisation, and freshness over live enterprise data. Together they cover the full spectrum of enterprise knowledge retrieval on GCP.
A retrieval-augmented generation system engineered for enterprises that cannot send proprietary data to public APIs. VPC-SC perimeter enforced. CMEK encryption — the enterprise holds the keys. Auditable retrieval chains with source citations on every response. The enterprise security argument for on-premise RAG made architecturally concrete.
Enterprise knowledge locked away because sending it to public LLM APIs is a compliance violation
VPC-SC perimeter · CMEK · private Vertex AI endpoint · Firestore audit log
Vertex AI · VPC-SC · CMEK · Cloud Run · Firestore · Gemini
Decomposes complex enterprise questions into optimal sub-queries, executes them in parallel across indexed corpora, re-ranks results by relevance and recency, and synthesises a coherent answer with citations. For questions too complex for single-shot retrieval.
Real-time RAG over live enterprise data — ERP records, CRM pipeline, ticketing systems, live inventory. Pub/Sub maintains a continuously updated vector index. Every retrieved fact carries a "last updated" citation and a staleness confidence score. For questions that need today's answer, not last week's document.
Three LLM engineering projects that go beyond prompting — compressing foundation models into domain experts, aligning them to enterprise safety constraints, and merging specialised models into unified capability. Raw intelligence, shaped for purpose.
Fine-tuning and distillation pipeline that adapts a foundation model to enterprise-specific vocabulary, contract language, and compliance terminology — producing a smaller, faster, domain-expert model deployable on-premise. The business case for not running a 70B model when a 7B model that actually knows your domain is measurably better.
Generic LLMs hallucinate on domain-specific enterprise vocabulary and regulatory terminology
Knowledge distillation · domain fine-tuning · on-premise deployment · model compression
Vertex AI · Hugging Face · PyTorch · Vertex Model Registry · Cloud Run
Fine-tuning a foundation model using RLHF and preference data to align it to enterprise safety constraints and refusal behaviour — for regulated industry deployments where "just prompt it" is not sufficient. The internal assistant that never hallucinates a dosage, always cites its source, and refuses off-topic queries by design. Model Cards with red-teaming results included.
Model merging using SLERP and TIES-merging to combine two specialised fine-tuned models — a contract-language expert and a financial-reporting expert — into a single model with both capabilities, without catastrophic forgetting. One internal assistant. Multiple domains. No fleet of models to manage.