We're an AI engineering
practice. Not an agency.
Agencies pitch. We build. The distinction is real: every engagement starts with a scope, carries an eval harness, ships behind a CI gate, and ends with you owning 100% of the code and infrastructure.
A 5th, AI CompMate, is in active development for our first enterprise client.
AI consulting has a
quality problem.
Most AI engagements end with a demo, a Notion doc of prompts, and a handover meeting. The client owns a prototype that works in controlled conditions and breaks under production load. No eval harness, no CI gate, no observability, a ChatGPT wrapper with a nicer interface.
We set up as an engineering practice because the problem was never "AI isn't good enough", it was that nobody was building it like engineers. Eval harnesses. Regression gates. System boundaries. Read-only architecture. Observability on every run.
The systems on our work page are the proof: TAFI 279/279, Orions 55/55, VistaGPT 95.3/100. Not demos, production systems on real workflows, every day.
One system.
Then four.
TAFI was the first, the AI layer on TechAppForce, Tech Extensor's metadata-driven low-code platform. A pipeline that takes a plain-English description and produces a complete, registered application with a passing QA suite. 279 eval stories, 100% pass, ~76% less code than the version before it.
The eval-harness pattern that worked for TAFI became the standard for Orions (incident diagnosis) and VistaGPT (ERP intelligence, five construction clients in production). Different domain, same foundation: define the output, write the stories, ship the gate.
AI CompMate, our organisational AI OS, is the fifth, in active development for our first enterprise client. Same standard, applied to a system still being built rather than one already running.
Agentix Tech was formed as the architecture-first AI consultancy arm of Tech Extensor, to take that engineering pattern to clients with repeatable specialist processes and no clear path to automation, the ones who needed rigour, not prompt engineering and good intentions.
TAFI, Orions and VistaGPT run in production today; AI CompMate is in active development for our first enterprise client.
- Chamberlin
- Delnor
- EmerySapp
- JC Construction
- JCWIP
- StubsAI
- A&A
- Cameron / Elite Capital
Not a startup. The AI arm
of a software company.
Agentix Tech is a subsidiary of Tech Extensor, the company our founder has built and led for over a decade.
Tech Extensor is an Ahmedabad-based software company with a 100+ person team and a serious moat in Healthcare IT, FHIR, HL7 and HIPAA-compliant systems, built over more than ten years. TechAppForce and its AI layer TAFI, the systems on our work page, were architected inside that group.
Agentix exists to take that engineering standard to clients as an architecture-first AI consultancy. So when you engage us, you're not betting on a two-person startup, you get a small, senior team backed by a company that has shipped regulated enterprise software for a decade.
"Systems should observe and remove work, not add it. Depth over hype, clarity over chaos.", Bhavik Thakkar, Founder
- 100+engineers in the parent group
- 10+ yrsbuilding enterprise software
- Healthcare ITFHIR · HL7 · HIPAA-compliant
- TAF · TAFIlow-code platform + AI layer, in-house
Lean on purpose.
Built to move.
How a small team out-ships bigger ones: fewer people, more ownership, and AI doing the grind. This is the operating model, not a wish-list.
Small team, by design.
Five to ten people, max. We scale with AI, not headcount, if we can't multiply ourselves with AI, we can't credibly do it for you.
Builders, not managers.
No managers managing managers. Everyone delivers, founders included. Get the people right and most "policy" stops being necessary.
Speed is the currency.
What legacy firms couldn't do in twenty years, new names shipped in under one. That pace is the bar, and why we move on new models and tooling early.
The loop is the method.
Think hard once, up front. Set the objective, set the guardrails, then let execution run in parallel. We design the system that does the work, not grind every task by hand.
Own your front.
Individual ownership, end to end. Full flexibility and every resource we can give, in exchange for sincerity and outcomes. That trust is the whole deal.
Talk less. Build more. Own your front.
Seven engineers.
One standard.
Everyone here has shipped production code into at least one live system, founders included. Nobody has only "advised" on AI.
Builder at heart, 15+ years in .NET and enterprise engineering. Owns and leads Tech Extensor Pvt Ltd, the 100+ person software company, and founded Agentix as its AI arm. Stakeholder in TechAppForce and its AI layer, TAFI.
Ships into TAF · TAFI →
The core technical builder behind every Agentix system. Designs the complex AI/ML architectures and automation, and architected TAFI, Orions and VistaGPT across ~3,953 commits, leading every technical engagement from the keyboard.
Ships into TAFI · Orions · VistaGPT →
Builder of the central knowledge base every Agentix AI service reads from, a retrieval-augmented (RAG) layer with vector search, and author of the VistaGPT MCP server that exposes it as tools and context to agents. Also ships production automations: cold-email and LinkedIn outreach and a YouTube Shorts autopilot, plus work across AI CompMate.
Ships into VistaGPT · MCP · AI CompMate →
Built TAFI's dynamic multi-agent orchestration, the planning, routing and tool-calling that coordinates its agents, along with its RAG knowledge bases. Specialises in conversational voice AI: real-time speech-to-text → LLM → text-to-speech pipelines on ElevenLabs, Vapi, Retell and Bolna, with deep automation across sales and marketing tools.
Ships into TAFI orchestration →
Builds complex agentic workflows on VistaGPT, embedding tool-calling LLM agents into legacy and multi-SaaS ERP systems with structured validation at every step. Owns VistaGPT's Excel automations, turning spreadsheet-bound finance and operations processes into governed AI workflows.
Ships into VistaGPT workflows →
Author of VistaGPT's MCP servers and the agentic SDK that runs agents against complex ERP systems, a dynamic execution runtime built on context engineering that assembles precise, scoped context and governs how the knowledge base flows into each agent. Owns VistaGPT's Power BI automations and the AI solutions around them.
Ships into VistaGPT · Power BI →
Applied-research engineer on TAFI, where he built the dynamic MCP builder and owns cost optimisation across the stack, model routing, token and context budgeting, and prompt caching. Continuously evaluates new models, tools and techniques and integrates them into production, the R&D function whose exploration shapes key technical decisions.
Ships into TAFI MCP builder →Five principles,
non-negotiable.
Not aspirational values, actual constraints that determine what we build and how.
Eval before production, always.
No system ships without an eval harness: 100+ scenario stories, a CI gate on every PR. "It works in testing" isn't enough, it has to hold under regression on every future change.
Read-only by architecture, not by policy.
Systems read data. Writes need explicit human approval or a named, reviewed exception, enforced at the framework level, not by trusting that a prompt will hold.
You own everything. No exceptions.
Every repo, every infra config, every model integration. No managed service that can change its pricing, no proprietary platform between you and your system.
Complexity earns its place.
Every agent, vector store and scheduler is there because we couldn't solve the problem without it. Start simple, add complexity only when simple fails, document why.
Evidence-backed output only.
Every claim a system makes is cited to a specific record. Synthesis without attribution doesn't ship; deterministic validation runs before any result reaches a user.
Eight non-negotiables.
Every system, every engagement.
The criteria a system must meet before we call it complete. Not "where applicable."
- Eval harness, 100+ stories before production
- CI regression gate on every PR
- MCP system boundary, all external access typed
- Langfuse traces on every agent run
- Read-only access enforced structurally
- Per-tenant isolation verified by integration tests
- Shadow-mode deployment for pipeline upgrades
- Full codebase + infrastructure handover on day one
Why "not an agency"
isn't just positioning.
The difference shows up in the deliverable, the ownership, and what happens after the invoice is paid.
- Pitches frameworks and approaches
- Delivers prototypes and demos
- Charges a monthly retainer for "ongoing support"
- System stops working when they stop being paid
- You don't own the prompt templates
- Scopes against a defined output and eval criteria
- Ships production systems with CI gates
- Project engagement with clear deliverables
- System runs independently after handover
- You own 100% of the code and infrastructure
Ready to build
production-grade?
A 30-minute call. We'll tell you what the system looks like, what the eval harness covers, and what production deployment involves, an engineering discussion from the first conversation.