AI agent developer
I'm Abhishek Gajera, an AI agent developer for startups and businesses. I build agents that take real actions in your systems: answering customers with live order data, drafting purchase orders from sales, searching people the way a recruiter thinks. They're built with tool calling, memory and retrieval, and they stay safe with guardrails and human approval where it matters.
What an agent can take off your plate
A good agent replaces a repetitive job a person does today, inside the tools they already use. These are the kinds I build.
Answer questions with live data from your orders, inventory and delivery systems, and hand over to a person when they should.
Watch sales, stock and invoices, then draft the next purchase order or task for someone to approve.
Turn a plain-language request into a precise search over your own data, using semantic and structured retrieval together.
Handle calls and chats in a contact centre, keep the conversation's context, and log it without leaking personal data.
Read invoices, resumes and forms, pull out the fields you need, and check them against a schema before they reach your system.
Streaming assistants your own users talk to, with multi-turn memory and access to the parts of your app they're allowed to touch.
Proof
Not demos. These run for real businesses.
A streaming AI assistant for a multi-product commerce platform in Chile. It keeps multi-turn memory and uses tool calling to work with live orders, inventory and delivery, so answers come from the business's real data.
Reads POS sales velocity and supplier invoices, then drafts purchase orders. Nothing goes out until a person approves it: human in the loop by design.
An LLM reads each recruiter search, then a FastAPI engine runs hybrid retrieval over Weaviate, mixing semantic embeddings with structured scoring. The model sits behind a provider-agnostic layer with retries, circuit breakers and fallbacks.
Bots on Amazon Lex V2 and Amazon Connect with Python Lambda fulfilment, conversation state in PostgreSQL and PII-safe logging.
Repository-level engineering tasks with reproducible failures and objective tests, used to train and evaluate autonomous coding agents. The same discipline goes into testing the agents I build for clients.
How I build them
Most AI agents fail on the parts around the model. This is what goes into every one I ship.
We pick the task, the systems it touches and what "done right" looks like, using real examples from your business.
Typed tool calls into your APIs and database, with the least access the job needs.
Retrieval-augmented generation over your documents and records: pgvector, Weaviate or Qdrant, with hybrid search and reranking.
Schema-validated outputs, PII masking, and a person approving anything that spends money or changes something important.
Prompt versioning, regression tests against saved examples, and monitoring with Sentry and New Relic, so you can see what it did and why.
Stack
Questions
An AI agent developer builds software where a language model decides which steps to take and then takes them: looking things up, calling your APIs, drafting documents, updating records. The work is less about the prompt and more about the system around it: which tools the agent may use, what it remembers, how it is checked, when a person has to approve, and how you can see what it did.
OpenAI and Anthropic Claude for most work, Azure OpenAI where a client needs it, and open models such as Llama or Mistral through Ollama when data has to stay in-house. LangChain where it helps, plain Python or Node.js where it does not. I keep the model behind a provider-agnostic layer so it can be swapped later.
Yes. An agent is only as useful as the systems it can reach, so most of the work is connecting it to what you already run: your database, CRM, inventory, ticketing or email, through their APIs. If something has no API, we look at what can be read or written safely another way.
By limiting what it can do and checking what it does. Agents get only the tools they need, outputs are validated against a schema, sensitive data such as PII is masked, and anything that spends money or changes something important waits for a person to approve it. Every run is logged, and prompt changes are tested against saved examples before they ship.
It depends on how many systems the agent touches and how much of its work needs review. The usual path is a small, scoped first version that does one job well, measured against real examples, and then extending it. Tell me what you want it to do and I will give you a clear estimate.
Tell me which task eats your team's time and which systems it touches. I'll tell you whether an agent fits, and how I'd build it.