Hire an AI integration developer for features that ship, not demos
Soham Dang builds AI features into production systems — OpenAI API integration, MCP (Model Context Protocol) servers and workflow automation — with permission checks between the model and the data it can reach.
The gap between an AI demo and an AI feature is almost entirely about the boundary. A demo queries the database and shows the answer. A feature has to decide whether this user, in this tenant, is allowed to see that row — before the model does.
Most of the engineering in production AI lives in that layer, not in the prompt.
What has been shipped
An AI meeting assistant
Transcribes live meetings, generates structured minutes, then creates and assigns action items to owners over WhatsApp — removing manual note-taking and follow-up.
MCP servers in production
Task, lead, HR and finance data exposed through a tool layer that checks permissions before the model sees anything.
An on-platform assistant
Available across the modules of a 25+ module business platform, scoped to what the signed-in user can already access.
Bilingual voice and read-aloud
English and Hindi speech input and output in an education product, built on OpenAI APIs.
The rules that make it safe
The model never gets raw database access. It gets tools, and each tool applies the same tenancy and role checks the rest of the application does. If a user cannot open a record in the UI, the assistant cannot read it either.
Beyond that: deterministic fallbacks when a call fails or times out, no user-visible surface that depends on a model returning valid JSON on the first attempt, and logging that makes it possible to work out afterwards what the model was actually asked.
Workflow automation
Not everything needs a model. A good proportion of what gets asked for as "AI" is a workflow — a trigger, a couple of conditions and an action — and is better served by n8n automation with a deterministic path.
Being willing to say that is part of the job. Model calls cost money per invocation and fail non-deterministically; a rule that runs the same way every time is usually the better engineering answer.
Common questions
What AI work has Soham Dang shipped to production?
An AI meeting assistant on the OpenAI API that transcribes live meetings, generates structured minutes and assigns action items over WhatsApp; MCP servers exposing task, lead, HR and finance data through a permission-checked tool layer; and an on-platform assistant across a 25+ module business platform.
What is MCP (Model Context Protocol)?
MCP is a protocol for exposing data and actions to a language model as structured tools rather than raw access. In practice it lets you put an authorisation layer between the model and the system, so the model can only ever act within the permissions of the user it is acting for.
Can AI features be added to an existing application?
Yes, and that is usually the right order — the permission model and data access layer already exist, so the AI layer becomes a tool surface over them rather than a parallel system with its own idea of who can see what.
Does every AI feature need a language model?
No. A significant share of AI requests are really workflow automation — a trigger, conditions and an action — which n8n or plain application code handles more cheaply and more predictably than a model call.