Hoshang Sheth logoHoshang Sheth
Service Deep Dive

AI Automation & Agentic Systems.

I build AI-powered systems that automate repetitive, knowledge-heavy, and multi-step business workflows.

At a glance
  • Timeline · 5-12 weeks
  • Investment · ₹35,000 - ₹90,000
  • Model · Fixed scope, fixed price
  • Ownership · Full source, no lock-in
Recurring
the work keeps happening after you stop thinking about it
24/7
runs at 3am without anyone watching
₹1.5-8K
typical monthly AI running cost for an SMB workload
Swipe
Business Applications

One capability, four workflows.

Different business problems, same underlying engineering discipline. Here's where this typically gets applied.

Sales & Lead Operations

Leads qualified, enriched, and followed up on before they go cold.

Lead capture, qualification, enrichment, CRM updates, and follow-up drafting running continuously, instead of depending on someone being at their desk at the right moment. A lead who messages at 11pm on a Sunday gets the same instant, structured response as one at 11am on a Monday.

  • Lead qualification pipeline running 24 hours a day
  • Automatic enrichment from public and internal data
  • CRM auto-update on trigger events such as form fills, payments, and status changes
  • Follow-up drafts ready for review, not sent blind
₹35,000 - ₹70,0005-8 weeks
See it in action

Customer Support

Answers from your documents, with sources, and a clean handoff when it should not answer.

Support and knowledge agents that retrieve from your SOPs, policies, and product docs, respond in your context, and escalate to a human the moment they are out of depth. A customer gets a sourced answer in seconds instead of a "someone will get back to you" ticket.

  • Source-cited answers pulled from your actual documents
  • Works where your team already is, in Slack, Teams, WhatsApp, or a web console
  • Escalation with full context, not a dead end
  • Access control on what each user is allowed to see
₹55,000 - ₹90,0008-12 weeks
See it in action

Document Processing

Invoices, forms, and contracts extracted, validated, and routed without manual re-entry.

Extraction pulls the data out, validation catches what doesn't match, and the result lands directly in your CRM, ERP, or database, with anything uncertain flagged for a human rather than silently guessed. An invoice that used to take fifteen minutes of manual entry lands in your system before you've finished your coffee.

  • Invoice capture and ERP entry automation
  • Schema and field validation before anything is trusted downstream
  • Exception queue for anything the system is not confident about
  • Full audit trail of what ran, when, and what it did with each item
₹35,000 - ₹70,0005-8 weeks
See it in action

Internal Operations Agent

Reasons over your operational data and completes the task directly, not just a recommendation.

An agentic system that reasons over your operational data and takes action through the tools you already use, whether that is reordering, ticket creation, calendar booking or status updates, completing the task itself instead of suggesting what someone should do next. Restocking, opening a ticket, booking a slot: it happens while you're in a different meeting, not sitting on a to-do list waiting for you.

  • Multi-step reasoning across multiple internal sources
  • Tool-calling into your existing stack via API to actually execute the task
  • Human-in-the-loop confirmation for anything irreversible
  • Full observability into what the agent did and why
₹45,000 - ₹90,0006-12 weeks
See it in action
Swipe
Why This, Not an Agency or a No-Code Consultant

What's different about how I build.

Custom code where it matters, platforms where they’re enough

A simple integration tool when the workflow is straightforward, and fully custom code when there's real business logic, edge cases, or scale involved, so you're never stuck paying ongoing per-run fees for something that never needed to be that complex.

Full source-code ownership

Everything lives in your own repository from week one. No lock-in, and no dependency on me to keep it running after handover.

One senior engineer, not a rotating team

You talk to the person actually building your system, every time, not an account manager relaying to whoever's free.

AI only where it earns its place

Some workflows need an agent. Others just need structured output and deterministic code. I use the simplest architecture that reliably solves the problem, not the most impressive one.

Swipe
The Problem

Why this matters right now.

Manual work quietly compounds. Leads qualified by hand at 11pm. Invoices copy-pasted between systems. The same question answered for the fifth time because the answer lives in a document nobody can find. Every hour spent on these is an hour not spent growing the business, and each one is a candidate for an AI-powered system built around how the workflow actually runs.

The Approach

How I build this.

  1. 1We walk through the workflow together: every step, every exception, every place a person currently has to make a decision. Then we agree on what should be automated and what should stay in human hands.
  2. 2I pick whichever approach the workflow actually needs, a straightforward integration tool for well-defined connections, or fully custom code when there is real business logic, multi-step reasoning, or scale involved. Nothing is automated blindly.
  3. 3Where the system needs to know something, whether that is your documents, SOPs, or product data, that knowledge is indexed and retrieved with sources, not left to the model to invent.
  4. 4Every system ships with a full audit trail and an exception queue, so you can see exactly what ran, when, and what it did with each item. Nothing fails silently.
  5. 5It connects to the tools you already use, such as your CRM, WhatsApp, email, or ERP, without replacing them, and goes live with monitoring in place plus full documentation and a walkthrough before handover.
What you get

Deliverables

  • Lead qualification and enrichment that runs 24/7 and updates your CRM without anyone touching it
  • Support and knowledge agents that answer from your documents with sources, and escalate cleanly when they should not answer
  • Document processing pipelines that extract, validate, and route invoices, forms, and contracts into your systems
  • Internal operations agents that reason over your data and take action through the tools you already use
Under the hood

Technology stack

PythonFastAPILangChain / LangGraphn8nRAG & Vector DBsMCPPostgreSQLStructured Outputs

n8n for well-defined integrations, custom Python and LangGraph for anything with business logic, edge cases, or multi-step reasoning. The architecture is chosen per workflow, not picked in advance.

Case Study

LeadPilot: qualified leads without staffing a 24/7 desk

Own Practice · hoshangsheth.com
Problem

Every visitor who wants to talk clicks through to WhatsApp, but a one-person practice cannot staff round-the-clock qualification. Quick questions and tire-kickers were taking the same attention as real, ready-to-buy leads, and anything that came in outside working hours just sat unanswered until morning.

Solution

LeadPilot receives the message, runs a structured conversation, and scores the lead against a transparent rule set, only escalating once someone is genuinely qualified, with a booking link and a notification sent automatically. It has been live in production since August 2026, tested extensively on real WhatsApp traffic, with nine real issues caught and fixed through that live testing, and running cost verified at roughly ₹1.34 per full qualifying conversation. It is the same system answering right now if you message the WhatsApp icon on this site.

Swipe
FastAPIPostgreSQLGemini 3.6 FlashMeta WhatsApp Cloud API
· Live in ProductionLive
Investment

₹35,000 - ₹90,000

Fixed scope, fixed price

Priced against the manual work it replaces and how much of the workflow it needs to cover. A system running a sales pipeline or a support desk unattended is worth more than its build effort. See the application breakdown above for where a typical project in each category lands.

Every project is quoted as a single fixed price once scoped. The range above reflects how much project size varies - not variable billing within a project.

Typical Delivery
5-12 weeks
Payment Terms
50% to begin · 20% at prototype · 30% before handover
Swipe
Frequently Asked

Questions specific to ai automation & agentic systems.

Depends on the workflow. For simple, well-defined connections I'll use n8n as it's faster to build and cheaper to run. For anything with edge cases, business logic, multi-step reasoning, or real scale, I write custom code so you're not paying per-run fees forever.
Every system ships with an exception queue and a human-review UI. Nothing is ever silently dropped, exceptions are surfaced with full context so someone can decide.
Yes. Standard integrations for Tally, Zoho, HubSpot, Salesforce, WhatsApp Business, Gmail, Slack, and most ERPs and knowledge tools with a REST API. Anything not on the list, I build a connector.
Most of what I build lives inside the tools you already use. A webhook here, an API integration there, so the output just shows up correctly in your CRM or ERP without anyone learning a new screen. A standalone interface only gets built when there genuinely is no existing place for the output to live, or when a human needs to review something before it goes out. We work out which one your workflow needs on the discovery call, not before.
Anything factual, pricing, specs, dates, quantities, is pulled from your structured data or retrieved from your documents with a source attached, rather than left to the model to invent. Anything uncertain is flagged for human review rather than guessed.
Every system ships with a monitoring view showing what ran, when, and what it processed. If something fails or hits an edge case, you get an alert and it lands in a human-review queue, nothing fails silently.
You own the code, the generated assets, and any prompts. LLM inference goes through your own API keys, and you can rotate providers at any time.
Every system ships with documentation and a config layer so your team can adjust rules without a code change. For structural changes, we do a short scoped update.
No. Every system runs on your own API keys, so your data goes to the LLM provider for inference only, not for training. If you need fully air-gapped deployment with no data leaving your infrastructure, that's a scope conversation we can have on the discovery call.
Next Step

Ready to scope ai automation & agentic systems for your business?

Bring the workflow you want to automate. We'll map it together on the call and I'll tell you exactly what can be handed off to a system.