Case studies from client and internal projects on Salesforce, Microsoft Azure, and applied AI. Client and product names are changed or left out.
Two Salesforce apps that turn photographed drinks menus and expense receipts into CRM records. A vision model reads the document, deterministic code checks and matches it, and nothing is written until a person has reviewed it.
Every store visit gets a Perfect Store score: configurable checks, points per threshold, and a traffic light. I built the calculation as an Azure Data Factory and Azure SQL pipeline that reads from Salesforce and writes the results back reliably.
Sales users wanted their Power BI reports directly on Salesforce records, filtered to the customer they're looking at. I built an embedding where Apex brokers the Power BI token, so users never sign in twice and no secret reaches the browser.
Field reps have targets such as visits, distributed products, or custom KPIs. I moved the calculation of their actual values to Python Azure Functions and Azure SQL, orchestrated by Azure Data Factory, and wrote the results back to Salesforce.
When a field rep closes an in-store order, Salesforce now generates a branded PDF with both signatures and emails it to the right people. Getting it to actually arrive meant diagnosing SPF, DKIM, and DMARC.
A German voice assistant for field sales got noticeably worse after 10–20 minutes of use. As part of the project team, I worked on tracing the problem through production logs to a browser speech-state bug and on making the agent's CRM writes verifiable.
Further client and internal projects, in short.
Set up a desktop AI system with vLLM behind an OpenAI-compatible gateway and measured throughput, tool calling, and fine-tuning, reaching 255 tokens/s at 100 parallel requests. A QLoRA fine-tune lowered accuracy, so I recommended the base model.
Backtested a risk model that decides which store visits get extra preparation time. With the same 20% capacity, it reached about 597 more at-risk visits than random selection (precision 88.5% vs. a 49.4% base rate, ROC-AUC 0.813).
Replaced a manually maintained Excel promotion calendar with a pipeline from the Fabric lakehouse through Power Automate to a VBA generator. It builds the calendar workbook automatically every weekday at 3:30.
Rebuilt a demo org from a read-only QA org: curated metadata, dependency-aware seed data, and 39,108 products pseudonymized and bulk-loaded. Guardrails refuse any write to the QA org.
Sales reps generate contract PDFs from a record in one click. A single orchestrator picks the template by contract type, so a new type needs one template and one mapping entry, with no change to the Flow.
A Screen Flow that deletes prospects already sent to SAP together with their contacts, visits, listings, and cases, or transfers them to another customer. Every write has a fault path, and any conflict rolls back everything.
Contacts linked to a retailer's head office appear automatically on all of its outlets. Bulk-safe Apex triggers keep the mirror in sync on insert, update, and delete.
A metadata-driven Apex batch creates holiday entries per field rep, sales org, and year, so they show up in the mobile app. It includes a correction for managed-package automation that reassigned the owners.
A command-line tool where one AI agent writes project documentation from a Git snapshot and a second agent checks every claim against evidence before anything is published.