Miran miransec

Miran

AI Engineer · Secure AI Systems · Backend · RAG

I build production-oriented AI systems where models operate inside reliable backend, retrieval, authorization, and security boundaries.

`~/currently`

shipped Averqen — AI-assisted security investigation with deterministic detection, FORCE RLS, grounded AI analysis, and policy-controlled response
shipped AtlasCore UI v2 — secure enterprise AI infrastructure for knowledge, retrieval, and grounded AI workflows
shipped VaaniDesk — multilingual AI customer-support platform with controlled actions and hybrid RAG
exploring the intersection of AI engineering and cybersecurity

Featured systems

Averqen

AI-assisted security investigation with deterministic security boundaries.

A security platform where telemetry ingestion, detection, correlation, tenant isolation, and response policy remain fully deterministic — AI operates only on bounded evidence with citations, never as the authorization or execution boundary.

Python · FastAPI · PostgreSQL FORCE RLS · Redis · Next.js · grounded AI · MITRE ATT&CK · adversarial testing

v1.0.0 verified:

1,493 tests passed · 0 failed
27 FORCE-RLS tables — tenant isolation enforced at DB level
58 OpenAPI paths — fully documented surface
Deterministic detection, correlation, entity graph, and threat intelligence
Grounded AI investigator with evidence citations, abstention, and MITRE ATT&CK mapping
Policy-controlled simulated response with human approval gates
MFA, scoped API keys, Redis rate limiting, SSRF defenses, hash-chained audit records

Repository: miransec/averqen

AtlasCore

Secure enterprise AI infrastructure for knowledge, retrieval, and grounded AI workflows.

A multi-tenant AI platform built around database-enforced isolation, workspace-scoped knowledge, hybrid retrieval, evidence-gated answering, and auditable access control.

FastAPI · PostgreSQL RLS · pgvector · Redis · Next.js · RBAC · grounded AI · evaluation

UI v2 verified:

717 backend tests passed · 0 failed
46/46 deterministic evaluations passed
FORCE RLS and restricted runtime DB-role invariants preserved
Hybrid FTS + pgvector retrieval with Reciprocal Rank Fusion
Evidence-first answering with citation validation and abstention

Repository: miransec/atlascore

VaaniDesk

Multilingual AI Customer Support Platform

A production-oriented support system for multilingual conversations, controlled business actions, access-controlled knowledge retrieval, and secure customer workflows.

multilingual AI · controlled tool calling · human approval · hybrid RAG · pgvector · authorization · idempotency · prompt-injection defenses

v1.0.1 verified:

206 backend tests passed · 0 failed
113 deterministic evaluations · 40 security-critical · 0 security failures
14 Playwright E2E tests passed

Repository: miransec/vaanidesk

`~/stack`

AI / ML
Python RAG Embeddings LLM APIs Evaluation

Backend
FastAPI PostgreSQL SQLAlchemy Redis REST APIs

AI infrastructure
pgvector Docker OpenTelemetry

Security
FORCE RLS RBAC adversarial testing audit logging

Frontend
Next.js TypeScript Tailwind CSS

Engineering
Git GitHub Actions pytest Ruff mypy

`~/principles`

models are untrusted components
authorization belongs outside prompts
retrieval should be measurable and evidence-backed
sensitive actions require explicit approval
state-changing tools should be idempotent
production claims require tests and evidence

`~/identity`

class Miran:
    focus = [
        "AI systems",
        "backend engineering",
        "secure retrieval",
        "agent security",
        "production infrastructure",
    ]

    shipped = ["Averqen", "AtlasCore UI v2", "VaaniDesk"]

    def philosophy(self) -> str:
        return "Build it. Test it. Measure it. Secure it."

Navigation Menu