✉ rajkumar@codeeasylabs.com ☎ +91 99001 20071
You're viewing the technical deep-dive on GMMCode — architecture, governance and proof. Prefer the short version? Talk to us and we'll walk you through it.
GMMCode · Enterprise Beta

The AI coding platform that plans before it writes, and governs every line it ships

Any LLM can generate the code: Claude, Gemini, GLM 5.2, or your own model. A locked plan and a governed inference mesh of 37 fine-tuned SLMs decide whether that code is ever allowed to reach your repository.

ITIL L1/L2/L3 Front Door Nine-Role Delivery Pods 5 Hard Closure Gates 37 Governance SLMs 302 Enterprise Standards Any LLM Provider Self-Hosted Option
governance.log · step 1 of 4 GMMCode beta
Developer: Add a JWT refresh-token endpoint to the banking auth service
⬠ GMMCode · Phase A: Planning
codebase_reviewer: scanned auth module, found 3 related files
GLM 5.2: drafted DevelopmentPlan, 4 steps, LLD attached
Open questions: 1, "Refresh token TTL: 7 days or 30 days?"
📋 Plan locked after your answer, no code written yet
⬠ Phase B: Execute step 1 of 4
Inject-first: 16 SLM checklists sent to GLM before it writes
GLM 5.2: generates code against all 16 checklists at once
🔍 Parallel review (16 SLMs, concurrent):
✓ security_scanner JWT validation correct, RS256 enforced
✓ exception_handler custom hierarchy applied
⚠ test_quality_guard missing expiry-assertion test
Convergence: 15 / 16 PASS (94%), above 80% threshold, revision applied, step complete

Governance log from a beta run. Every step produces a log like this: which SLMs ran, what they flagged, how the plan was revised.

3
Service tiers (L1/L2/L3)
in front of every request
9
Delivery roles per pod
coordinating by recorded mail
5
Hard closure gates
with segregation of duties
37
Governance SLMs in the mesh,
up to 16 active per generation
302
Enterprise coding standards
across 16 domains
43
Enterprise languages
across 7 industry tiers
What Is GMMCode

Provider-agnostic code generation. Non-negotiable governance.

GMMCode is CodeEasy Innovation Labs' governed AI coding platform. Any large language model can handle code generation: Claude, Gemini, GLM 5.2, or a future in-house model. A separate layer of fine-tuned Small Language Models enforces your coding standards, security policy, and compliance requirements on every task, automatically.

The core invariant behind the product: the coding provider is swappable. The governance is not optional.

  • A locked plan and Low-Level Design before a single line of code is written
  • 37 governance SLMs in the mesh, up to 16 reviewing each generation in parallel, not after the fact
  • Runs on your own GPU endpoint, your code and standards never leave your environment

See the Full Governance Model

GMMCode, by CodeEasy.

Platform Architecture

Three layers between a request and a release

A user request never goes straight to a model. It passes through service management before any code work begins, and code work passes through governance gates before anything closes.

REQUEST   A user reports an issue or asks for a change
1

ITIL Service Front Door: L1 / L2 / L3

Three support tiers own user contact, triage, troubleshooting, and code-level investigation before anything reaches a developer, human or AI. Tickets, SLA clocks, a Known Error database, and a knowledge base sit underneath.

Single point of contactTool-enforced escalationINC / SR / PRB / CHG
2

Nine-Role Delivery Pod

A full delivery team of role agents, coordinating only through recorded mail, activated event-driven by unread messages, watched by a stall watchdog. Closure is hard-blocked until five gates carry approved verdicts.

9 roles5 hard gatesRecorded-mail coordinationHumans + agents
3

Governed Inference Mesh

Deterministic retrieval matches each code chunk to the standards that apply. Up to 16 of the mesh's 37 fine-tuned domain SLMs validate concurrently. Conflicting verdicts are arbitrated with a fail-closed instruction: when in doubt, the code does not ship.

37 SLMs in mesh302 standards16 domainsFail-closed arbitration
RELEASE   Code reaches your repository only with five approved gate verdicts and a full governance log
Email Intake

Work starts in the inbox, not the editor

Anyone on the allowlist raises work by email. No prompt courses, no new tools, no onboarding decks. Adoption cost is zero because the behaviour already exists.

Zero training

To creates the ticket. Cc adds context. Attachments become evidence. Your team already knows how to use GMMCode; they have been practising every working day.

Easy in, governed out

A two-line email lands in a strict system: L1 triage, priority from impact and urgency, a nine-role pod, five hard gates. The user supplies intent. GMMCode supplies rigour.

🏢

The whole company drives it

Competitors sell to developers. GMMCode is driven by sales, finance, operations, anyone with an inbox. Only developers ever see the IDE.

Threaded replies update the same ticket. Sender allowlist per engagement. Own-mail loop prevention. A full governance log on every step.

Service Front Door

L1 → L2 → L3: service management in front of every request

Most AI coding tools hand every prompt straight to a model. GMMCode puts a full ITIL service layer in front, so known issues are answered instantly, service is restored before code is touched, and code changes only happen through the governed pod.

L1

Service Desk: Single Point of Contact

Owns every user conversation. Resolves from known sources only: open and closed tickets, the Known Error database, and the knowledge base.

  • Categorises incident vs service request
  • Derives priority from impact × urgency
  • Registers the ticket and its event trail
  • Only role holding respond_to_user, one voice to the user
L2

Engineer: Restore Service First

Reads logs, runs diagnostics, applies or invents workarounds. Resolving without leaving a knowledge-base article behind is treated as unfinished work.

  • Raises problem records for recurring incidents
  • Maintains the Known Error database
  • Classifies code fixes as bug, change, or new feature
  • Escalates to L3 with evidence attached
L3

Investigator: Code-Level Analysis Only

Reads the codebase, correlates log evidence to code paths, and writes a technical brief: suspected root cause, reproduction, blast radius, acceptance hints.

  • Raises a linked change record
  • Hands off to the delivery pod with the brief as its task
  • Never edits code, the service layer cannot bypass the gates

Tier boundaries are code-checked, not prompted: L1 attempting to read logs, or L3 attempting a knowledge-base search, is rejected with a tier-boundary error. Every escalation automatically updates the ticket's assigned tier, status, and event trail.

Pod Execution

A nine-role delivery pod behind every change

Code work is not done by one agent with one prompt. It is delivered by a full pod of role agents, each with its own bounded context, its own inbox, and its own accountability.

Product Owner Business Analyst Architect Developer Tester Security Engineer Code Reviewer DevOps Engineer Delivery Manager

The Product Owner closes the work item, and deliberately owns no gate, so the closer can never self-certify.

Closure is hard-blocked until five gates carry approved verdicts

Design Gate
Signable only by the Architect
QA Gate
Signable only by the Tester
Security Gate
Signable only by the Security Engineer
Code Review Gate
Signable only by the Code Reviewer
Release Readiness Gate
Signable only by the DevOps Engineer
📧
Recorded-mail coordination
Roles coordinate only through a cloud mailbox with to/cc/read-by tracking. Whoever has unread mail acts next, and every handoff is on the record.
Stall watchdog
If the queue empties without closure, the Delivery Manager is alerted automatically. Stuck work escalates to a human instead of silently dying.
🤝
Hybrid human + agent teams
Any role can be held by a human or an agent. The gate table treats their verdicts identically, and refuses signatures from the wrong side, keeping segregation honest both ways.
📁
Per-engagement isolation
Every project carries its own mailbox, plan, status log, and gate history. One long-lived engagement carries many concurrent work items, contexts never mix.
How It Works

Two phases: Plan, then Execute

Most AI coding tools write code the moment you ask. GMMCode plans first, locks that plan with you, and only then writes code, one step at a time, under governance.

Phase A: Plan

A codebase-review SLM scans the relevant files. GLM 5.2 drafts a full DevelopmentPlan with a Low-Level Design and any open questions, using PMP and LLD templates injected as constraints, not left to chance.

Confirm: Lock the Plan

You review the plan, answer any open questions, and lock it. No code exists yet. This is the alignment gate that most AI coding tools skip entirely.

Phase B: Execute, Step by Step

Each step is generated one at a time. Every governance SLM injects its checklist into the prompt before GLM writes, then validates the result concurrently. A step only completes at 80%+ convergence.

Under the Hood

What happens when you give GMMCode a task

Task received
Your prompt arrives in the IDE. The codebase-review SLM reads the files it touches before any plan is drafted.
📋
Plan drafted
GLM 5.2 produces a DevelopmentPlan and Low-Level Design, using injected PMP and LLD templates. Open questions are surfaced instead of guessed at.
🔒
Plan locked
You confirm the plan, with clarifications if needed. This is the alignment gate, nothing is generated before this point.
🔨
Inject-first step execution
Each locked step is sent to GLM one at a time, with all relevant governance SLM checklists injected into the prompt upfront. The model is asked to self-enforce before it writes.
🔍
Parallel governance review
Up to 16 of the mesh's 37 review SLMs (security, exceptions, logging, concurrency, tests, and more) validate the generated code concurrently, not sequentially.
Convergence and drift check
At 80%+ SLM pass rate the step is marked complete; below that, GLM revises and the review runs again. A scope guard confirms the result still matches the locked acceptance criteria before moving to the next step.
Proof

Numbers from the governance log

Every GMMCode step writes a governance log: which SLMs ran, what they flagged, how the plan was revised. The figures below come from that telemetry.

94%

Median first-pass convergence

Share of governance SLM verdicts passed on the first generation of a step, before revision.

1 in 6

Steps revised before shipping

Generations below the 80% convergence threshold, revised and re-reviewed before completing.

<4 min

Prompt to locked plan

Median time from task prompt to a confirmed DevelopmentPlan with LLD and resolved open questions.

GMMCode is running structured 30/60/90 day pilots with design partners, including a New Zealand enterprise engagement and delivery teams in Bengaluru. Case studies publish during the beta, with partner names where permitted.

What You Get

Everything the GMMCode beta ships with today

🔬

Governed Inference Mesh

37 fine-tuned governance SLMs, up to 16 running in parallel on every generation (security, compliance, architecture, tests), before code reaches your repo.

📋

PM-Gated Planning

Creates a locked DevelopmentPlan with an LLD and open questions before writing a single line of code. No more unreviewed blobs.

🧩

Built on VS Code

All your extensions, settings, themes, and keyboard shortcuts work out of the box. GMMCode is a governed layer, not a rewrite.

🔒

Zero-Trust Code Gen

Nothing ships without passing the governance threshold, checked as part of generation, not as a post-hoc linter.

💬

Multi-Turn Memory

Full conversation history with file-context injection. GMMCode tracks your codebase, it does not hallucinate missing imports.

📤

Export + Audit Trail

Export any chat session as Markdown. A full governance log per step, which SLMs ran, what they flagged, how the plan was revised.

How We Compare

GMMCode vs standard AI coding tools

Other tools generate code the moment you ask. GMMCode plans it, locks the plan with you, then governs every line as it is written.

Capability GMMCode GitHub Copilot Cursor Claude Code
Plan locked with the user before any code is written EnforcedNoOptionalPlan mode, not enforced
ITIL service front door (L1/L2/L3, tickets, SLA)NoNoNo
Multi-role delivery pod with hard closure gates 9 roles, 5 gatesNoNoSubagents, no gates
Parallel multi-model governance review at generation time Up to 16 SLMsNoSingle-model reviewSingle-model review
Enforced enterprise coding standards library 302 standardsNoNoNo
Compliance evidence trail (SOX / GDPR / HIPAA / RBI / DPDP)NoNoNo
Fully self-hosted stack, including models (DPDP Act 2023)NoNoVia cloud VPC only
Segregation of duties enforced in codeNoNoNo

Comparison reflects publicly documented capabilities as of July 2026. All four tools support multiple model providers. The differentiation is governance.

See Standards, Compliance & Security Detail

Roadmap

From governed IDE to governed delivery organisation

GMMCode is infrastructure for autonomous software engineering. The IDE is the entry point. The destination is a delivery organisation where humans and agents share one audit trail.

Today · Beta

The Governed IDE

PM-Gated Planning, the governed inference mesh, and a full audit trail on every generation. Delivered to design partners on macOS, Linux, and Windows through structured enterprise pilots.

Next

Autonomous Pods on Your SDLC

Nine-role delivery pods running whole work items end to end inside your service management framework, with humans holding whichever roles you choose.

Then

The Governed Delivery Organisation

Humans and agents in one audit trail, across hundreds of concurrent engagements, with every decision reconstructable after the fact.

Why This Exists

Built by someone who lived inside the audit

Rajkumar Neelappa

Rajkumar Neelappa · Founder & CTO, CodeEasy Innovation Labs

Rajkumar spent 22 years delivering enterprise software across Siemens, IBM, P&G, HP, and Capgemini, across four countries. Change advisory boards, segregation of duties, release gates, audit evidence: he enforced them by hand for two decades, then watched them break the moment AI code generation arrived.

GMMCode puts that discipline inside the pipeline. The controls run in code, so ungoverned output has no path to a repository.

BE, National Institute of Technology Karnataka, Surathkal.

SiemensIBMP&GHPCapgeminiNIT Karnataka, Surathkal

Implemented with your team. Full governance features from day one.

GMMCode is deployed through a structured 30/60/90 day enterprise pilot: scoped on a real engagement, implemented alongside our platform team, and validated against your own standards and service management framework. PM-Gated Planning, the full SLM review pipeline, and every governance control are included, with no feature gating. Design partners lock in preferred commercial terms and shape the roadmap.

FAQ

Common questions about GMMCode

Is GMMCode the same product as CodeEasy?
Yes. GMMCode is the customer-facing brand for CodeEasy Innovation Labs' governed AI coding platform. "CodeEasy" remains the internal and technical name used across our source code, APIs, and documentation, it is the same underlying product, not a separate tool.
How many governance SLMs actually run on my code?
The governed inference mesh holds 37 fine-tuned domain SLMs covering 302 standards across 16 domains. Deterministic retrieval matches each code chunk to the standards that apply, which activates up to 16 SLMs concurrently per generation. You never pay latency for validators that are irrelevant to the code in front of them.
What happens when a user reports an issue, does it go straight to an AI developer?
No. Every request enters through the ITIL service front door. L1 answers from known sources, tickets, the Known Error database, the knowledge base. If nothing is found, L2 restores service with diagnostics and workarounds. Only when a code fix is genuinely required does L3 investigate, write a technical brief, and hand off to the delivery pod through a linked change record. L3 never edits code, so the service layer can never become a bypass around the gates.
How does pod execution with roles actually work?
Each work item is delivered by a nine-role pod: Product Owner, Business Analyst, Architect, Developer, Tester, Security Engineer, Code Reviewer, DevOps Engineer, and Delivery Manager. Roles coordinate only through recorded mail; whoever has unread mail acts next. Closure is hard-blocked in code until five gates (design, QA, security, code review, release readiness) carry approved verdicts, each signable only by its owning role. Any role can be held by a human or an agent; the gate table treats both identically.
What is PM-Gated Planning, and why does it matter?
Before any code is generated, GMMCode reads the relevant part of your codebase and drafts a full DevelopmentPlan with a Low-Level Design and any open questions. You confirm or clarify that plan before a single line of code exists. Most AI coding tools skip this step entirely, which is exactly where uncontrolled, unreviewed code tends to come from.
Does my code or governance data leave my machine?
No. The governance SLM layer runs on your own private endpoint, and each client's custom governance model is stored in an isolated, encrypted namespace. For organisations with DPDP Act 2023 data-residency requirements, the full stack, including the coding model, can be self-hosted on Indian-resident infrastructure.
Which AI models does GMMCode work with?
Any provider. The current beta routes coding generation through GLM 5.2, with Claude and Gemini available as alternate providers. The governance layer is separate from the coding model by design, you can swap the coding provider without losing any governance capability.
How much does GMMCode cost?
GMMCode is deployed through structured enterprise pilots during the beta, with no licence fee for pilot partners and no feature gating. Pricing after beta will be per developer seat, with a self-hosted enterprise option. Pilot partners lock in preferred rates ahead of general availability.

Plan it. Lock it. Let GMMCode govern the rest.

PM-Gated Planning, a governed inference mesh of 37 SLMs, and 302 enterprise coding standards. Implemented with your team through a structured enterprise pilot.

Request Enterprise Demo