Engagement Brief — Full DetailEngagement

Enterprise AI & MLOps Modernization

Two production engagements delivered for a leading AI computing company · Azure-native · Confidential · 2026

Engagement 01 · Enterprise AI · RAG Consolidation

Unified Agentic AI Platform

Seven siloed RAG agents — each built on Python and Microsoft Autogen — consolidated into a single agentic platform. An intelligent MCP router and Agent-to-Agent (A2A) protocol give users one conversational surface across Teams, Slack, and Web, with every existing backend preserved behind lightweight adapters — no rewrites, no regression risk.

Unified Experience

One bot across Teams, Slack & Web via the M365 Agents SDK — natural-language access, no routing knowledge required.

Intelligent Orchestration

Dual-path MCP routing (rules + AI intent), A2A cross-agent handoffs, secure document orchestration.

Enterprise Foundation

Centralized Entra ID auth, unified state (Redis + Azure SQL), one telemetry view across every agent.

ChannelsTeams · Slack · Web

GatewayAPIM · Entra ID

MCP HostRouter · A2A

7 Python AgentsAdapters · unchanged

The Challenge

Seven separate applications — several on the now-deprecated Bot Framework — each with its own auth and context. Users had to know which tool to ask, and no agent could hand work to another.

Our Approach

An orchestration layer above the agents on the M365 Agents SDK. Zero backend change via adapters, centralized security at the gateway, and a progressive path from rule-based to ML routing.

PHASE 0

Discovery

Architecture & contracts

PHASE 1

Foundation

CIA & PDP pilot

PHASE 2

A2A

Cross-agent handoff

PHASE 3

Scale

All 7 agents live

PHASE 4

Optimize

ML intent routing

Engagement 02 · Silicon MLOps

OpsML — Test Automation & Modernization

The client's wafer and chip test-scoring pipelines ran on licensed, GUI-driven KNIME Server — costly, hard to scale, and locked in. Banjubits re-engineered them into containerized Python services on Azure Kubernetes Service, validating scoring behavior node-by-node while eliminating license dependency and adding modern engineering controls.

Automated Ingest

Test files parsed, validated & deduplicated, then merged into Azure SQL with newest-test-wins persistence.

ML Scoring + Rules

CatBoost, XGBoost & Random Forest inference, with DPAT and outlier rules applied after the models.

Resilient Ops

Flock-guarded scheduler, deadlock-retrying SQL, transaction audit trail, structured logs & alerts, CI/CD.

File ShareTest-data bus

AKS WorkersPer-product · cron

Azure SQLMERGE persist

ML + RulesScoring · DPAT

Foundry FilesBin-map output

The Challenge

Scoring pipelines ran on licensed, GUI-driven KNIME Server — expensive to renew, hard to version and scale, and locked to a single vendor's runtime.

Our Approach

Node-by-node re-engineering into containerized Python on AKS — validated to 100% scoring parity — with CI/CD, structured logging, and a full transaction audit trail.

Two engagements, one operating standard.

From enterprise AI to silicon MLOps, we modernize without disruption — preserving proven logic, running Azure-native, and shipping governance and observability from day one.