Tariq S.

Senior AI Engineer | RAG, AI Agents & SaaS | Python, FastAPI, Next.js

I build production RAG pipelines with measured retrieval accuracy, LLM agents with guardrails and evals, and I audit existing AI builds before they meet real users. Python, FastAPI, Next.js, OpenAI and Claude. Most of what I am handed already technically works on the founder's machine. The gap is everything between that and paying customers. For nine years, founders have handed me a product and trusted me to own the whole thing, from architecture through the AI layer to the frontend and deployment. Increasingly the work is not building from zero. It is fixing what a fast moving founder or an AI coding tool already shipped, and making it safe to put in front of real users. Recent builds: a healthcare compliance RAG agent on LangChain, GPT-4o and Pinecone, a multi tenant AI SaaS with tenant isolation enforced at the database level and Stripe billing, and a multi agent operations platform that reads from and writes to a CRM unattended. Most of my delivered work sits in healthcare and telehealth, fintech and financial operations, real estate, and compliance heavy operations where being wrong is expensive. AUDITS Often the first thing a client needs is not more code, it is an honest read on what they already have. I run paid technical audits on existing AI SaaS builds and hand back a written assessment: critical issues and technical debt, security and data handling gaps, reliability and error handling across webhooks and background jobs, database improvements, cost and latency waste in the AI layer, and a prioritised plan with effort estimates. On one healthcare compliance build the same approach took audit time down 85 percent and human error down 90 percent. If you want the same engineer to implement the plan afterwards, I do that too. REGULATED ENVIRONMENTS I am currently the founding tech lead on a HealthTech platform working with Epic and HIPAA workflows, and I have shipped telemedicine and financial platforms where access control and sensitive data handling were constraints from the start rather than a later patch. RETRIEVAL THAT HOLDS UP AS THE CORPUS GROWS Document parsing, semantic and sliding window chunking, embeddings, hybrid search with cross encoder re ranking, tenant and document level filtering so users only ever see what they are authorised to see, and grounded answers with source citations from OpenAI or Claude. Vector databases including pgvector, Qdrant and Pinecone depending on scale and budget. Most RAG systems fail on retrieval, not on the model. AGENTS THAT RUN UNATTENDED Tool and function calling, structured outputs, multi agent orchestration with LangGraph and LangChain, retries and clear stopping conditions, human in the loop checkpoints, cost ceilings, and monitoring so you find out an agent is misbehaving before your customers do. I have built agentic workflows that read from and write to CRMs unattended, and document pipelines where every model call passes a validation layer before anything is committed. Evaluation is part of the build, not an afterthought. Golden question sets, retrieval precision and answer groundedness scoring, and regression runs before deploys. I do not ship an AI feature I cannot measure, and when quality drops I can tell you whether it is a prompt problem, a retrieval problem or a data problem. THE REST OF THE STACK React, Next.js, TypeScript, Node.js, NestJS, Python, FastAPI, PostgreSQL, Supabase, Docker, AWS, Vercel, REST APIs, webhooks and streaming responses, Stripe, authentication and RBAC, CI/CD. HOW I WORK I am not an order taker. If you have a clear product vision I will argue with you about scope: what belongs in version one, and where your current architecture breaks under real users. Clients keep me long term for the boring reasons. Clear written updates, risks flagged early, honest estimates, clean documented code in your repo. I work US hours from Pakistan, five days a week: 8 AM to 6 PM US Eastern, 7 AM to 5 PM Central, mornings through early afternoon Pacific. Replies land inside your working day, not overnight. Certifications: AI Automation Engineer (Google, 2025), Full Stack AI Engineer (Coursera, 2024). The portfolio below has the specifics. Most engagements start small. A paid audit, or a single scoped milestone. Send me a short description of what you are building, or what is already broken, and I will reply with how I would approach it and what I would do first.