Sachintha
Antany.

SAP Business One AI Integration MCP Server for Gemini Enterprise

Neither SAP nor Google offers a native integration between SAP Business One and Gemini. I independently researched SAP's Service Layer OData API and designed a custom Model Context Protocol (MCP) server from scratch, transforming SAP Business One from a transactional repository into an autonomous cognitive engine. Rather than navigating dozens of nested SAP windows and static Crystal Reports, this architecture enables real-time natural language querying, cross-entity reconciliation, proactive anomaly detection, and automated document generation.

The server exposes 75 read-only MCP tools across Procurement, Sales & CRM, Financials & Accounting, Inventory & Assets, Contracts & Service, and System & Metadata. Gemini orchestrates multi-step operations across these tools: autonomous 3-way PO matching that cross-references purchase orders, GRPO receiving slips, and vendor invoices to flag quantity/price discrepancies automatically. Intelligent PR/PV generation where Gemini parses an uploaded vendor invoice PDF via multimodal ingestion, validates vendor codes and GL accounts against live SAP data, and produces a complete purchase requisition with zero manual data entry. Dynamic cash flow analysis combining bank statements, incoming payments, vendor payments, and open receivables to calculate rolling net cash position and flag liquidity shortfalls. Forensic inventory variance detection comparing physical counts against perpetual stock ledgers to identify shrinkage, dead stock over 90 days, and carrying cost impact. Customer margin intelligence computing gross profit per customer and flagging margin dilution from extended payment terms or heavy discounting.

Combined with Google Workspace, the AI cross-references SAP data with Drive spreadsheets, correlates Gmail threads with SAP transactions, and synthesizes answers from multiple sources into a single response. Executives get revenue summaries, margin health, and aged debt overviews in plain language without needing SAP desktop clients. Estimated 80%+ reduction in AP/procurement processing time and 15x faster daily data retrieval.

SAP Business One Conversational Agent on Gemini Enterprise

Business users had to open SAP Business One and walk through its screens to pull routine ERP data such as sales orders, invoices, inventory, debtor aging, and financials, so everyday questions were slow and dependent on whoever knew the system. I built a high-code Google ADK agent (Gemini 2.5 Pro) that lets anyone ask in plain language and get live, accurate figures back, including charts and forecasts. The agent reaches SAP through the custom MCP server (a separate project) over Streamable HTTP, exposing 80+ read-only Service Layer tools across sales, purchasing, payments, inventory, finance, banking, service, and master data, plus ARIMA time-series and MRR-based forecasting.

It is served as an A2A service and runs on Gemini Enterprise, but because it is a standard ADK/A2A agent it is portable to any surface that speaks A2A, not tied to one UI. It deploys to Vertex AI Agent Engine and reaches the Cloud Run MCP server in a different region through a managed Agent Gateway egress path, with a CA bundle wired in for TLS.

The hard engineering sits in three layers. A reworked MCP session manager fixes a real failure mode: when the MCP server's Cloud Run instance scales to zero the HTTP connection still looks healthy but server-side session state is gone, so the next call 404s on a session the stock manager keeps reusing. A session TTL plus explicit invalidation force a fresh session on a failed call. A lenient A2UI toolset parses, repairs (including bracket-recovering truncated JSON), auto-injects missing ids and form bindings, and validates against the Gemini Enterprise composite catalog before handing UI back to the client. A deterministic fast path renders interactive finance canvases from live SAP data with no LLM round trip: it detects greetings, menu phrases, dashboard triggers, and round-tripped canvas button events, carries a fiscal-year-aware period parser (fiscal year starts in April) covering quarters, month-name ranges, and "last N months", and uses a ping probe to tell a genuine SAP outage from other failures so it never hallucinates a dashboard.

DSI Route Optimizer: AI-Powered Delivery Route Planning for 111+ Showrooms

Sri Lanka's largest footwear retailer (D. Samson & Sons) operates 111+ showrooms across all 25 districts. Every morning, delivery managers manually plan truck routes from the Colombo warehouse to 10-25 showrooms along a corridor. Wrong sequencing burns hours backtracking through congested roads. I built an AI-powered route optimisation agent that solves this as a single-vehicle open-path Travelling Salesman Problem, deployed inside Gemini Enterprise where managers interact with it conversationally.

The agent uses Google OR-Tools with Guided Local Search to find near-optimal stop sequences for 25 locations in under a second, at zero API cost. Travel times come from Google's Routes API with TRAFFIC_AWARE routing, not straight-line approximations, capturing real Sri Lankan road conditions: single-lane coastal highways, mountain switchbacks, and Colombo urban congestion. A data pipeline discovered and verified all 111 DSI locations via Google Places API, auto-classifying each into districts and delivery corridors, stored in Firestore.

The agent emits structured payloads rendered as A2UI Material Design components natively in Gemini: selection chips, toggle controls, data tables, and expandable driver manifests, with zero frontend code. Managers can iteratively refine routes: remove stops, add detours, switch between shortest-time and shortest-distance. The open-path TSP formulation (distinct start/end depots) models DSI's actual operations where trucks end at their final destination rather than returning to origin. For a 15-stop run, even a 10% reduction in travel time saves 20-30 minutes daily across fuel costs, driver utilisation, and faster showroom restocking.

Google Billing to SAP Business One Middleware: Reconciliation & Invoice Processing

Each month Finetech receives Google billing (Workspace, Cloud Platform, and Maps) covering many customer domains, and finance staff reconciled each domain against the matching customer account in SAP Business One by hand, then keyed the result back into SAP. It was slow and error prone, so it kept getting deferred until a long backlog built up, and it is harder still because customers bill on different cycles (monthly, quarterly, annual), so a customer's SAP invoice need not fall in the same month as the Google invoice, which breaks any date-based match. The COO, Ms. Sujatha Nadesan, had wanted this solved for years, so I built a middleware that reads SAP directly, does the matching and pricing maths, keeps a human in the loop, and writes the result straight back into SAP.

The app is a modular monolith: a Python 3.12 FastAPI backend (services across 10 modules, roughly 27 endpoints) behind a React 19 and TypeScript single-page app (around 7,000 lines) built with Vite 6, Tailwind CSS 4, TanStack Query, Zustand, react-hook-form, and Zod. Two write-capable workflows run against SAP B1. Reconciliation is a header-driven parser that ingests Google Workspace invoices and GCP/Maps cost tables (.xlsx and .csv), groups lines by domain or billing account, resolves each to its SAP service-group vendor card by matching the domain against SAP's CardForeignName, classifies Workspace versus Cloud lines, and computes the amount to reconcile per service group. The non-obvious part is the billing-cycle handling: rather than filtering SAP invoices by the Google invoice's month (which would hide quarterly and annual customers), it pulls open invoices for the business partner and matches on amount, not date.

The second workflow, Invoice Processing, backs a New Request intake form that looks up the SAP customer card by domain and searches the SAP item master so staff never retype master data, with pricing and proration maths computed live in the front end (unit-tested with Vitest) and raised as an A/R or proforma invoice in SAP. Both workflows share one SAP Business One Service Layer session client, and every SAP write is funnelled through a single posting method gated by a server-side control, so live updates switch on deliberately after validation and finance sign-off rather than by accident. Access is Google OAuth (OpenID Connect) restricted to the company domain, layered over Cloud Run IAM, with server-side roles (finance, technical, admin) and an admin user-management UI. Per-user drafts save and resume to Firestore, every uploaded file is archived to a Cloud Storage audit bucket, and secrets live in Secret Manager. It ships through a multi-stage Docker build with a Cloud Build trigger auto-deploying separate pre-prod and production environments.

Executive Financial & Debtor Dashboard over SAP Business One

Executive financials (P&L, operating expenses, debtor aging, invoice listings) were maintained by hand in an Excel workbook fed from SAP exports, a workbook that still carried live #REF! errors, so leadership could be reading broken numbers without anyone knowing. There was no single, always-current, access-controlled view of company performance, and every refresh meant a manual re-export and re-paste. I built a standing web app that pulls the same figures live and read-only from SAP Business One and reconciles exactly to SAP's own Profit & Loss and Customer Receivables Aging.

A FastAPI backend (around 2,600 lines of Python) sits on the SAP B1 Service Layer via an async httpx client that manages the login session and issues read-only GETs, exposing five dashboard sections across seven endpoints. The P&L and Opex logic follows SAP's own account structure (the first digit of each GL account's FormatCode) rather than a hand-kept account map, nets credit minus debit per drawer and per two-digit opex sub-group per month, and caches the chart of accounts because a six-month pull is roughly 20,000 journal-entry lines. The reconciliation-grade calculations are the hard part: debtor aging is an exact as-of balance computed by a GL control-account roll-back (customer current balance minus post-as-of journal movement), bucketed into SAP's native aging intervals with GETs parallelised behind a concurrency-limited semaphore; the Invoice Listings view reproduces SAP's Crystal "Post Costing Summary" per document and reconciles it against the general ledger so the tiles sit on the same GL basis as SAP's P&L.

A React 18, Vite, and Tailwind front end (around 3,500 lines) served by FastAPI from the built bundle gives a collapsible sidebar shell over five sections (Overview, P&L, Opex, Debtor Aging, Invoice Listings), a fiscal-year-aware reporting-period selector, CSV export, and per-tile "how this is calculated" popups. Access runs through Google OAuth (OpenID Connect) with per-section RBAC, the user store in Firestore in production and a local JSON fallback in dev. It ships through a multi-stage Docker build auto-deployed by Cloud Build on push to main, with the SAP password mounted from Secret Manager and no write path compiled in. The result replaces an error-prone manual workbook with a live, access-controlled dashboard whose P&L and debtor aging reconcile to SAP to the cent, verified per customer and in total.

Finetech IDP: Internal Developer Platform for GCP

Researching and building an Internal Developer Platform that automates the provisioning of customer environments on Google Cloud. A non-technical operator selects a customer, a product, and a runtime, and the platform creates the infrastructure, deploys the application, and returns the service URL. Replaces a fully manual GCP provisioning process with a self-service portal backed by reusable Terraform modules.

The architecture enforces single-tenant isolation: every customer environment gets its own GCP project (ft-<customer>-<env>), designed for banks and enterprises with strict compliance requirements. A three-layer Terraform module library (Foundation, Runtime, Data) with strict contracts means adding a new runtime (Cloud Run, App Engine, GKE) is one new module directory with zero changes to products, API, or UI. Supports two hosting models from the same codebase: Finetech-hosted (project under Finetech's GCP org) and BYOC (customer's own GCP, accessed via Workload Identity Federation).

The provisioning flow runs through a plan-then-approve gate: Terraform init, plan, human approval with resource summary (+N ~N -N), then apply, with live logs streamed to the portal via SSE. Full RBAC (admin/operator/viewer), audit trail, and a product catalog with pluggable config schemas. Currently running AI products including a 360 evaluation agent and a Gemini agent bot on the platform.

Sela: trilingual AI shopping concierge for Kapruka 🛍️

A full-screen chat agent that takes a shopper from "I need something" all the way to a paid order with delivery, without ever leaving the conversation. It works for gifting and everyday shopping across Kapruka's real catalog, understands and replies in English, Sinhala, and Tamil, and accepts voice, transliteration typing, and handwriting input. An animated mascot reacts to every stage of the journey.

The entire storefront is the chat. Sela connects to Kapruka's Model Context Protocol for real product search, delivery quotes, a multi-item basket, gift messaging, and guest checkout with a live payment link, then streams tool results back as generative UI: product cards and carousels, delivery cards, and order cards, never walls of text. A language picker starts the chat in English, Sinhala, or Tamil and Sela mirrors the user's script, including Singlish and Thanglish. Users can speak, type Singlish or Thanglish that converts to native script as they type, or draw letters on a handwriting pad, plus a hands-free voice mode where Sela speaks back in the chosen language. Date awareness resolves relative dates like "next Friday" to real Sri Lanka calendar dates, and a resilient multi-model brain with automatic fallback (Gemini to Groq to Claude) keeps the agent up even when a provider is rate-limited.

Hive Trendz: AI social media automation

Autonomous news-to-post pipeline: discovers trending articles, generates branded content via LLaMA 3.3 70B (Groq API), and publishes to Facebook & Instagram with zero manual effort. Flutter app with on-device image compositing, FFmpeg video editor, and AlarmManager exact scheduling. Flask + APScheduler backend on Railway with a full web dashboard for queue management and live logs.