AI app development that holds up in production

What I build

• Multi-agent systems in LangGraph: a router, specialist agents, shared memory, and a QA agent that checks every answer before it goes out.
• MCP servers that give Claude, ChatGPT or Cursor narrow, safe access to your database or API.
• RAG and memory: Pinecone vector search for what the AI should remember, and a normal database for facts that must be exact.
• LLM features inside your product: drafts, summaries, classification and triage.
• Evaluation and guardrails: LLM-as-judge checks, result caps, read-only access, logging.
• AI in mobile apps: answers streamed to React Native over Server-Sent Events. (How: streaming LLM responses to React Native.)

Who this is for

• Teams with a working product who want one AI feature that actually gets used, not a chatbot bolted on the side.
• Companies whose answers are locked in a database only one person can query.
• Healthcare, finance and other fields where a wrong answer costs something, so checks and limits matter.
• Mobile apps that want an AI coach, assistant or search that feels instant.

AI work I’ve shipped

• Myaigi

  Seven specialised agents behind an intent router in LangGraph, across Claude, OpenAI and Gemini, with Pinecone memory and a judge agent gating every answer.

  Architect

• MCP server at Sehat Kahani

  A Python MCP server and Chainlit chat that let ops staff ask a production database questions in plain English.

  Built it alone

• SOAP note drafting

  An LLM feature that drafts structured clinical notes from patient consultations, to cut documentation time.

  Led development

• Baba Telehealth triage

  A fine-tuned OpenAI model for AI-assisted patient triage ahead of the consultation.

  Team delivery

• Open source

  A merged fix to the DeepSeek-Reasonix engine (PR #8198) and features for the PSX MCP server.

  Contributor

How an AI project runs

1. Pick one jobWe choose the single task the AI should do well and collect 20–50 real examples of it.
2. Prototype on your dataA working prototype against those real examples before any interface is built.
3. MeasureA small test set and judge checks, so “good enough to ship” is a number both of us agreed on.
4. Ship with guardrailsLimits, logging, fallbacks and a cost dashboard, then iterate on what users actually ask.

Pricing

Model usage is billed by the provider to your own account, so you always see the real running cost.

Why me

• I’ve shipped AI where mistakes matter: healthcare and personal coaching, with gates that stop bad answers.
• I build the product around the model: backend, cloud, web and the React Native app, not just the prompt.
• I understand what makes AI slow and expensive, down to prefill, decode and the KV cache. I wrote it up in LLM inference for app developers.

Questions clients ask

Can you add AI to our existing app or product?

Yes, that’s the most common request. I start from one job the AI should do well, prototype it against your real data, and only then build it into your backend, web app or React Native app.

Which AI models do you work with?

Claude, OpenAI and Gemini, often in the same system. On Myaigi a fast, cheap model routes requests and stronger models handle generation and review, so you don’t pay top prices for every call.

Is our data safe when an AI model uses it?

It should never hold your credentials. I give models narrow tools with typed inputs, keep access read-only unless writing is the point, cap result sizes, and keep secrets in a managed store. That’s how I’d build an MCP server over a production database today.

Should we use RAG or fine-tuning?

Start with retrieval (RAG) in most cases: it’s cheaper, you can update it daily, and you can see what the model was given. Fine-tuning earns its cost when you need a consistent format or behaviour at scale, as with the triage model on Baba Telehealth.

What does it cost to run AI features?

Build cost follows the pricing above. Running cost is the model provider’s bill, paid on your own account. I design to keep it low: smaller models where they’re good enough, caching, and short prompts.

Do you build MCP servers?

Yes. I built an MCP server that gives an LLM safe access to a production MySQL database, and I’ve contributed to an open-source MCP server for Pakistan Stock Exchange data.