A 3D portrait of George Mridun, smiling, wearing a cap and glasses.

George Mridun

AI solution architect, Melbourne, Australia

I find a real problem, research it, then design and build the AI solution on GCP, AWS and Azure.

Every figure comes from deterministic tools, never the model. Try RepoRecon and SiteRecon live.

RepoReconSiteRecon

RepoRecon: 78% precision and 88% recall on 16 hand-labelled fixtures

Experience

  1. Founder & AI Solution Architect, GEOSTACK

    2024 – PRESENT

    • Approach: start from a defined problem, research how it is handled today, then design the agent workflow, the checks around it and the cloud it runs on. Every solution ships with an evaluation.
    • Workflow: each system is a staged agent pipeline (RepoRecon six agents, trase seven roles, FinSight nine in parallel) with typed outputs between stages. Code decides and validates; the model drafts and explains.
    • Evaluation: every agent workflow is measured, not asserted. Hand-labelled benchmarks, golden cases, evidence gating, schema validation, rubric critic loops, prompt-injection tests and baseline comparisons, with the gaps stated where a build has no benchmark yet.
    • Shared eval layer across projects: logs every model call's output, latency and cost and scores answers against hand-written golden cases, so a prompt regression is caught before release.
    • Cost control: a trained logistic-regression classifier resolves confident RepoRecon findings in sub-millisecond time at $0; only low-confidence cases reach an LLM.
    • Cloud: RepoRecon and SiteRecon run on GCP (Compute Engine, Docker sandbox, Caddy, Workload Identity Federation, so no long-lived keys in CI). CareCert runs on Vercel and Supabase in Sydney. Azure and AWS experience comes from my HSW and Aus Biz roles.

  2. IT & Cloud Systems Engineer, Howards Storage World

    MAY 2025 – PRESENT

    • Introduced AI ticket triage on the IT queue, reducing daily tickets from about 20 to about 5.
    • Rolled out an internal policy Q&A assistant with source citations on every answer. To address colleagues' accuracy concerns, I demonstrated its failure modes side by side. The same document set is also served as an MCP server.
    • Built the data and automation layer: sales, payments, inventory and e-commerce data into Supabase and BigQuery on GitHub Actions schedules, SQL views for faster refresh, and a Microsoft Fabric warehouse over several source databases.

  3. Systems Analyst, Cloud & DevOps, Howards Storage World

    MAY 2024 – MAY 2025

    • Built middleware automation that syncs 26 store profiles between the store data source and Birdeye and the store locator.
    • Hardened Azure and Microsoft 365 identity and handled incident response.

  4. AWS DevOps Engineer (Internship), Aus Biz Consultancy

    NOV 2023 – MAY 2024

    Lambda, API Gateway, CloudWatch

  5. ML Data Associate, Amazon Development Centre

    SEP 2021 – JAN 2022

Projects

RepoRecon

Live · Multi-agent, security audit

Paste a public GitHub repo URL and a six-agent pipeline clones it into a network-isolated Docker sandbox, runs Semgrep and Trivy scans, and produces a plain-language security and architecture report. A public /accuracy page publishes the system's measured precision and recall against a labelled benchmark, including the weaker runs.

reporecon.georgemridun.dev

SiteRecon

Live · Agent, evidence-gated scoring

Crawls a site, its public social profiles and competitors, then scores SEO and AI-search readability. Every score change must cite a verbatim quote from the page, or the change is rejected.

siterecon.georgemridun.dev

CareCert

Live · Compliance copilot, NDIS

An AI-first compliance copilot for NDIS providers facing mandatory registration. Providers manage registers, incidents, workers, evidence and audit packs, with a chat copilot grounded in official guidance. A deterministic rules spine decides compliance; the model only explains.

ndiscert.vercel.app/

Carbonly (trase)

Built · Multi-agent, certificates

An assessment and support assistant for Australian environmental certificates (VEU and ESS). Seven role-separated agents sit behind a deterministic core that produces every figure.

Research Conductor (AskMyDocs)

Built · Research agent, RAG

Began as document Q&A and grew over seven versions into an autonomous research pipeline. A LangGraph orchestrator scrapes several academic databases in parallel, routes across models, then runs a Writer↔Critic loop until the draft passes a rubric.

FinSight

Built · Multi-agent, finance

A multi-agent investment research platform. Nine specialist agents with typed Pydantic outputs run in parallel and stream results to the UI over WebSockets.

Exilink

Live · AI matching, marketplace

A talent marketplace with an AI matching layer: an LLM parses resumes, skills and jobs are embedded, and Qdrant similarity ranks candidates for a role.

exilnk.vercel.app

ISM control mapper

Built · Fine-tuning, security

Fine-tunes a small open model on Kaggle's free 2×T4 GPUs to map a security finding to the Australian Information Security Manual control it breaches. The model predicts the control ID only.

Contact

For AI solution architecture, agent workflows and cloud delivery, get in touch.