Harkirat S.

AI Backend Architect | Full Stack AI Development | System Design

I build the backend AI products run on: APIs, data models, queues, retrieval and the infrastructure underneath. The interesting part of an AI product is rarely the prompt. It is what happens when a thousand requests arrive at once and something has to stay correct, fast and affordable. 🧰 What I build APIs in Python and Node that serve AI features under real concurrency Retrieval layers over your own data: ingestion, chunking, embeddings, indexes, reranking, cache Streaming responses, tool calling and agent loops with timeouts and limits that actually stop Background workers for the long jobs, with retries, dead letter queues and idempotency Event pipelines so every request leaves a trace you can audit later Auth, rate limits, quotas and tenant isolation, built in rather than bolted on afterwards 🧩 System design I get asked for Schema design for data that is part relational and part vector, without ending up with two sources of truth Choosing between one service and several, and usually arguing for fewer than you expected Caching strategy at each layer, which is normally the largest single cost reduction available Model routing and fallback, so one slow provider does not take the product down with it Read and write paths separated when analytics starts competing with the application Migration plans that run alongside the live system instead of requiring a night everyone dreads 🔒 The parts people skip Secrets, key rotation and least privilege access, done before an enterprise buyer asks Personal data mapped, minimised and deletable on request Prompt injection treated as untrusted input, because retrieved documents and user text are exactly that Cost ceilings and alerts per feature, so a runaway loop is caught in minutes Backups tested by restoring one, which is the only test that counts 🔧 How I work Read the code and the traffic before proposing anything Write the design down: the constraint, the options, the tradeoff, the decision Load test against realistic numbers so capacity is measured rather than assumed Ship in reversible steps behind flags Leave dashboards and runbooks, because on call at three in the morning is not the time to reverse engineer intent 📦 What you get Working services in your repository with tests around the parts that matter Infrastructure as code, environments and a deploy pipeline your team can run without me Documented API contracts your frontend or client teams can build against immediately Observability: metrics, traces and logs wired up, with alerts that mean something A capacity and cost note: what breaks first, at what volume, and what it costs at ten times today ⚡ Stack Languages: Python, TypeScript, Node, some Go where it fits Frameworks: FastAPI, Django, Express, NestJS Data: Postgres, pgvector, Redis, MongoDB, ClickHouse, Elasticsearch AI: OpenAI, Anthropic, open weight models, RAG pipelines, evaluation harnesses Infrastructure: AWS, GCP, Docker, Kubernetes, Terraform, GitHub Actions Async: Celery, BullMQ, Kafka, SQS, Temporal 🧭 How the first two weeks usually go Read the code, the schema and the traffic before offering an opinion Reproduce the actual problem under load rather than trusting the report of it Come back with the three things worth changing, in the order I would change them, with the cost of each Fix the cheapest high impact one first so you get a result while the plan is still being agreed You get the written analysis regardless of whether you continue with me 🤝 Working together Direct communication, no layers, no rewritten status updates I will disagree with a plan before it is built rather than after Estimates come with the assumptions that would change them Comfortable overlap with European and US hours, and reachable when something is on fire ✅ Good fit if You have or expect real traffic and correctness matters more than novelty Your AI feature works in demos and falls over in production You want someone who will own the design and then implement it You are willing to spend a little time on the boring layer so the product stops surprising you Tell me what is slow, expensive or fragile, and roughly what your traffic looks like. I will tell you where I would start and what I would leave alone.