Abdul Rafay Ahmed

Enriching a CRM meant a different vendor, export and spreadsheet for every field, and hand-written integration code that broke whenever a provider changed its API.

Provider connections over OAuth (PKCE, dynamic client registration) with encrypted token storage and silent refresh. Tool discovery over each provider's MCP server, so the planner reasons over live capabilities instead of hand-coded adapters. A plan generator on AWS Bedrock that turns the prompt plus Salesforce field metadata into a typed, multi-step plan with confidence, credit cost and latency estimates, dry-run against real records before any credits are spent. Execution as a streaming Node.js pipeline: mass runs over whole datasets, single-record runs, and real-time create and update triggers, with web search and page fetch available as tools, credit accounting, and write-back to Salesforce. Plus the Vue prompt editor, the folders view and the credits page.

Periodic clean-ups meant duplicates lived in the CRM for days before anyone saw them. Catching them live means three things can touch the same dataset at once: a change event, an automatic merge and a scheduled analysis.

On the Node.js side, the real-time path: Salesforce change events queued per dataset, automatic merging within confidence ranges with an AI review step, Salesforce Flow invocation when a match is found, and rollback for Person Account merges. In the Python engine: similarity matching with MinHash and locality-sensitive hashing, so fuzzy duplicates are found across 10M+ records without comparing every pair; an information-theoretic method (normalized mutual information) that picks which field combinations are worth comparing and excludes expensive ones; incremental analysis so only changed records are re-examined; and match-confidence breakdowns that show users why two records were paired.

Freelancers paid from abroad either pay an accountant they never meet or skip filing, drop off the active-taxpayer list and pay double withholding tax on everything from bank withdrawals to car registration.

A pure, versioned rules engine for the tax year with worked-example tests as the regression net. A wizard that parses payment-platform exports, computes the return live as the user types, and produces a filing pack. Magic-link auth and row-level security on Supabase, Paddle checkout plus a manual bank-transfer flow with entitlements enforced server-side, transactional email on Resend, static marketing pages and the app on Cloudflare Pages with an Express API on Workers.

Brand tracking is only as good as its inputs, and consumer Arabic is not the Arabic that off-the-shelf models were trained on. Egyptian, Gulf, Levantine and Maghrebi text was being mislabelled.

An async FastAPI orchestrator over Apify collectors for Facebook, Instagram, TikTok, X, Google Reviews and TripAdvisor, with de-duplication, per-platform normalization and Terraform-managed Azure infrastructure. Arabic NLP models (dialect, sentiment, emotion, irony) across 20+ dialects behind BrandHealth360, served on Redis-queued FastAPI workers, with training and deployment automated with Docker and Airflow. A multi-tenant RAG layer (Pinecone, Chroma, LangChain agents) with structured-output guardrails feeding dashboards and CRM integrations.

A live broadcast has to cut to the right player within moments of an elimination. A director watching seven screens cannot do it reliably, and there was no API to ask.

Seven game feeds captured in parallel, OCR on each feed to detect elimination events as they appeared on screen, elimination tracking across the bracket, and the integration that switched the broadcast to the eliminating player's feed and webcam. Tested against recorded matches, then run live.