zabdalabs

Zabda Labs — Rengarajan G · AI Systems, Independently Built and Verified
CLAIMS POLICY — every statement on this page is verifiable: demoable, runnable, or filed.
BENGALURU, IN · ZABDA LABS
Rengarajan G · AI Engineer · Independent Builder

AI systems that can be checked, not just believed.

Ten years running e-commerce operations. Two years building AI infrastructure independently — governance engines, agent-state protocols, and on-device inference. Three patent applications filed. One product live with paying users. Everything below is stated at the level I can demonstrate.

portfolio_audit.log2026
$ tsc --noEmit · vnol-core0 errors
$ vitest run · vnol-core44 passed
$ tsc --noEmit · sentrix0 errors
$ cargo test · chaytpassing
cikitsu · productionlive + billing
patent filings · IPO India3 filed
phantominfer · benchmarksresearch
# run on real hardware, June 2026
WHY THE STAMPS

I build verification tools, so this page is written like one. Each project carries the status an auditor would give it — nothing is stamped verified unless it compiles, its tests pass, and it runs today. Codebases are private while patent filings are active; architecture walk-throughs and live demos are available on request.

Project ledger

06 flagship entries · status as independently checked
01

Sentrix

AI execution integrity — patent filed (India)

The problem

Teams deploy AI tools with no control over what the AI actually does. It can drift from the task it was given, breach policy, or act outside its mandate — with no record of when or how.

What it does

Locks the user's intent at the start of a task, then checks every AI output against it in real time — flagging deviation, drift, and policy violations with a full audit trail. Ships as a CLI, VSCode and Chrome extensions, and an MCP server.

How it differs

Most guardrail products inspect content — toxicity, PII, unsafe topics. Sentrix checks fidelity to intent: did the AI do what was actually asked. Those are complementary layers, not substitutes.

Honestly: intent-integrity is a young category and larger security vendors are now entering it. My edge is shipping depth — working IDE, browser, and MCP integrations today — not incumbency.

TypeScriptVSCode APIChrome ExtensionMCPNode.js
Verified · Runs
~6.9k lines core TS
0 compile errors
CLI + clients working
02

VNOL

Agent-state portability protocol — patent pending

The problem

When an AI agent crashes, a session ends, or you switch providers, its working state — goals, reasoning, constraints, next steps — is gone. Teams restart complex work from zero and stay locked to one vendor.

What it does

Serializes an agent's full working state into a compact, AES-256-GCM encrypted, integrity-sealed snapshot — and restores it on a different model (OpenAI, Anthropic, Llama, Mistral) so work resumes where it stopped.

How it differs

Framework checkpointing (e.g. LangGraph persistence) saves workflow state inside one framework. VNOL saves the agent's cognitive state as a model-agnostic, sealed artifact designed to boot on a different provider — with a security audit run before anything is sealed.

Honestly: if you live entirely inside one framework and one provider, its built-in persistence may be all you need. VNOL is for when you can't afford that lock-in.

TypeScriptnpmMCPAES-256-GCMMulti-model
Verified · Tested
7.4k lines core TS
44/44 tests passing
0 compile errors
03

Cikitsu

Multi-agent business strategy SaaS — in production

The problem

Small and mid-size businesses can't afford a cross-functional consulting team — finance, marketing, operations — working one problem together.

What it does

Eleven specialised AI consultants collaborate on a submitted business question — challenging each other's reasoning — and deliver a structured strategy report. Deployed with subscription billing and PDF reporting.

How it differs

It isn't a wrapper around one chat model: agents form teams dynamically and debate before concluding — and the whole product runs on the same governance and state layer (VNOL, UTCD) I build as infrastructure. It's the working proof of the stack.

Honestly: AI consulting tools are a busy space. Cikitsu's claim isn't "the only one" — it's "live, billing, and built on infrastructure I can show you."

Next.jsFastAPIPythonPostgreSQLDocker
Live · Paying users
Production deploy
Subscription billing
Built on VNOL + UTCD
04

Azhwar

AI code governance engine — patent pending

The problem

AI coding assistants generate code fast, and much of it quietly violates accessibility, security, or design standards — unnoticed until it breaks in production.

What it does

Audits AI-generated code at the syntax-tree level against 250+ rules across 30+ industry profiles, suggests corrections, and runs wherever code is written — npm CLI, GitHub Action, or MCP server inside the editor.

How it differs

General linters check syntax and style. Azhwar is profiled for the specific failure modes of AI-generated code, mapped to industry compliance profiles rather than generic rules.

Honestly: established scanners (Snyk, Semgrep) are strong on classic vulnerabilities. Azhwar's lane is the AI-slop layer they weren't designed for — it complements them rather than replacing them.

TypeScriptASTnpm CLIGitHub ActionsMCP
Verified · Ships
CLI + Action working
250+ rules
30+ profiles
05

Baliza

Bounded AI delegation for founders — launch ready

The problem

Founders can't hand communication to AI without guardrails — it may approve things they wouldn't, miss their tone, or overstep authority they never granted.

What it does

The founder defines authority rules, escalation triggers, and voice first; the AI drafts only inside those bounds. Every action is logged, explainable, and reversible with a kill switch. Billing and onboarding built.

How it differs

Most email-AI products optimise for more automation. Baliza optimises for safe delegation — the rules engine and audit trail are the product, the drafting is downstream of them.

Honestly: it is launch-ready, not launched. The engineering is done; the market test hasn't started.

Next.jsNode.jsPostgreSQLGmail APIStripe
Launch ready
Phases 1–6 complete
Billing integrated
Pre-launch
06

PhantomInfer

On-device LLM inference engine — research

The problem

Most AI requires a cloud connection and sends data off-device — unusable in low-connectivity environments and risky in privacy-sensitive sectors.

What it does

A C++ inference engine for running language models directly on mid-range Android hardware — custom kernel, mixture-of-experts routing, thermal management, and JNI integration. Groundwork for private, offline AI.

How it differs

Rather than wrapping an existing runtime, it's a ground-up engine targeting the constraint nobody optimises for: 6GB-RAM consumer phones, not flagship hardware.

Honestly: this is research-grade. On-device throughput for large models on mid-range hardware is measured in tokens per second, not magic. I publish what the hardware actually does.

C++Android NDKCMakeJNIGGML
Research
70+ C++ source files
Custom MoE router
Active development

How I work

the rules this portfolio is written under
Claims match evidence

A number appears here only if I can reproduce it in front of you — a passing test suite, a clean compile, a live deployment. Where something is early, it says research; where it's untested in market, it says launch-ready, not launched.

Comparisons cut both ways

Every "how it differs" section also states where the alternative wins. If a built-in framework feature solves your problem, you should use it. The projects here exist for the cases where it doesn't.

One stack, not fifteen demos

The flagship products share infrastructure — the state protocol, the contract layer, the governance engine. Cikitsu runs on it in production. The portfolio is one system seen from different angles, not a pile of prototypes.

Extended stack

supporting infrastructure · built, private, demoable

Chayt

Rust · builds, tests pass

Governed memory gateway: AES-256-GCM encrypted memory, Ed25519 signing, contract enforcement at the proxy layer, Prometheus metrics.

UTCD

Python · packaged

A contract standard for AI agent behaviour — signed, validated tool and conduct descriptors. The protocol layer the other systems enforce.

cSIM

Prototype · patent material

Portable, hardware-attested AI cognitive state — compact enough to carry as a QR code. Encryption and secret-sharing core implemented.

Alocaka

Python · in development

Pre-execution governance for agent organisations: budget enforcement, policy checks, and pre-flight simulation before an agent acts.

aot-engine

Python · framework

Self-assembling multi-agent topology with dynamic leader election and fault absorption. Powers Cikitsu's consultant teams.

crazyant

TypeScript · CLI + MCP

Takes AI-generated prototypes to production: hybrid AST + LLM security audit, hardening middleware, infrastructure generation.

About

I spent ten years running e-commerce operations — supply chains, vendors, P&L. In 2024 I started building AI systems independently: no team, no funding, no classroom. Two years later the result is a connected stack of governance, state, and inference infrastructure, three patent applications, and one product in production with paying users.

The operations decade is why everything here is built around control, audit trails, and accountability: I've been the person responsible when a system silently does the wrong thing. The tools I build exist so that doesn't happen with AI.

B.Tech in Information Technology · Scaler Data Science certification · Bengaluru, India.

NAMERengarajan G
BRANDZabda Labs
PATENTS3 applications filed (India)
IN PRODUCTION1 SaaS, paying users
LANGUAGESTypeScript · Python · C++ · Rust
SURFACE AREAnpm · MCP · VSCode · Chrome · GitHub Actions
BASEBengaluru, IN

Verify it yourself.

Codebases are private while patent filings are active — but every status stamp on this page can be demonstrated live: test runs, deployments, architecture walk-throughs. If you're hiring, piloting, or evaluating, ask for the demo that matters to you.