Maksud Ul A.

I help CPA firms and fintech founders build AI tax planning platforms, AI agents and workflow automation, without an unreviewed number ever reaching a client. The AI reads documents and proposes. A deterministic engine does every calculation. A professional approves before anything goes out. That control structure is the difference between software your staff trust in March and software they quietly abandon. "We want our own AI tax planning platform for our clients." I have built one. FortunAI is an AI tax planning SaaS for a US CPA firm. It extracts data from 1040, 1120-S, 1065 and K-1 returns and financial statements, analyzes each client's position, applies the firm's own tax strategies to find savings, and produces a tax plan the CPA reviews and approves. Lara, an AI tax-expert agent, works inside it. It is multi-tenant, so the firm serves every client from one system. I led the 22-person team that delivered it, from requirements and Figma sign-off through architecture, sprints, code review, QA and DevOps. "We tried AI and it invented a number." The model never gets the chance to hallucinate a figure. It proposes strategies by ID from a library your tax team controls, and a rules engine performs every calculation with inputs logged. Anything outside the library is rejected. Field maps for returns are versioned by form and tax year, because line numbers move between years, and every extracted figure traces to its source by form, page and line. "Our AI agent works in the demo, then falls apart with real users." Production AI engineering, not demos. AI agents and multi-agent workflows on LangGraph and the Claude Agent SDK, with a bounded agent loop: every tool call validated, steps capped, failures retried or escalated to a human. RAG grounded in your own documents with citations, so answers trace to a source instead of hallucinating. An eval harness that scores accuracy and hallucination rate against real cases before every release. In production, AI DevOps: prompt and model versioning, tracing, guardrails, and token-cost monitoring, with model routing that sends simple steps to cheaper models. "We want to sell this to other firms." I design for that from the first sprint: firm-level tenant isolation, roles and permissions, audit trails, and a strategy library your tax team updates without a developer. Catalist, a multi-tenant company credit and ownership reporting SaaS my team built on the Equifax and CreditorWatch APIs, is a second example. "Our staff lose hours to email, document chasing and follow-ups." Workflow automation inside the firm: email triage and drafting, document requests, task routing, deadline tracking and status reporting, with an approval gate on anything that leaves the firm. Integrations: QuickBooks and Xero APIs, SendGrid, Azure Blob and AWS S3 storage, Azure Document Intelligence and Textract, n8n and LangGraph. "Taxpayer data cannot leak between clients." Encryption in transit and at rest, role-based access, no client data used for model training, and an audit log of every automated action and who approved it. "We need a CTO, not a full-time hire." Fractional CTO: roadmap, architecture ownership, team leadership, vendor and cost decisions. Currently Fractional CTO for an EdTech SaaS platform. ALSO BUILT CommEngine: AI-powered social CRM unifying Facebook, WhatsApp and Instagram conversations. Co-founder and CTO, $325K raised. Summit Alliance Port Limited: supply chain system running daily operations across three Chittagong terminals, a river terminal and a port in Kolkata. C#, Angular, PostgreSQL, on-premises. Fortune 500 delivery: John Hancock, Manulife, Unilever, Schiphol Airport, Netherlands Immigration. PROOF 20 years · 59 Upwork contracts · 5,400+ hours · 100% Job Success · Top Rated Plus 5 SaaS products built · Founder of one product that exited at $5M+ MSc CSE · Microsoft Certified Professional · Claude Code 101 and Building with the Claude API (Anthropic) · Azure AI-102 and AZ-305 in progress STACK AI engineering: RAG, AI agents, multi-agent systems, agent loops, eval harness, hallucination guardrails, LLMOps / AI DevOps, model routing, MCP AI: LangChain, LangGraph, RAG, Gemini, Claude, Azure OpenAI, Azure Document Intelligence, Textract, n8n Build: Next.js, Nest.js, Python/FastAPI, React, Angular, C#/.NET, PostgreSQL, Elasticsearch, AWS, Azure Tell me what you want to build or where your team loses the most hours. You will get an honest read on scope, cost and what to do first, including whether I am the right fit. Prefer to start small? A fixed-price two-week discovery with a costed MVP plan.