Agentic AI for legacy-to-cloud data platform modernization. An agent pipeline migrates enterprises off legacy systems and onto any modern cloud warehouse — with a human approving every AI-generated change.
Status: Working prototype / demo-stage · mock data · solo build · AIBoomi Startup Weekend
Live demo: https://jemathew.github.io/data-platform-modernization-agent/ · Demo video: {DEMO_VIDEO} · Repo: https://github.com/JEMathew/data-platform-modernization-
Why now. Legacy data platforms (Oracle, Teradata, SQL Server, Hadoop, Informatica) carry fixed cost, can't feed modern AI/analytics, don't scale, depend on a shrinking PL/SQL talent pool, and reach end-of-life — so enterprises must modernize. But ~83% of migrations run over budget or fail, with legacy complexity the #1 cause. A project is usually kicked off by a concrete trigger: a license renewal, an AI mandate, an end-of-life deadline, a capacity wall, M&A, or a stalled prior attempt — and each is a buying moment.
{PRODUCT_NAME} is a unified agent console that runs a pipeline of AI agents — Profiler, Mapper, Code-gen, Validator — to assess a legacy estate, map it to a modern target (Snowflake, BigQuery, Databricks, Fabric), generate the migration code, and validate the result. A deterministic engine does the schema translation; an optional LLM handles the ambiguous procedural code — and a human approves every draft before anything ships.
End-to-end automated migration already exists — but as services-heavy, six-figure accelerator engagements from specialist vendors. Frontier LLMs now make the hardest part — procedural-code translation — automatable as a self-serve product instead of a consulting project. Our wedge is the product-led experience and the in-app human-in-the-loop approval gate, not "vendor-neutral end-to-end" (which is now table stakes).
Why modernize (drivers). Legacy platforms carry fixed licensing/hardware cost, can't feed modern AI/analytics, don't scale, depend on a shrinking PL/SQL talent pool, slow time-to-insight, and reach end-of-life. Cloud warehouses/lakehouses fix all of these — but ~83% of migrations run over budget or fail, with legacy complexity the #1 cause.
What starts a project now (triggers). A license renewal or hardware refresh, an AI mandate, a cloud-first program, product end-of-life, a performance/capacity wall, M&A consolidation, a regulatory change, attrition of the last engineer who knows the legacy procs, FinOps cost-cutting, or a stalled prior attempt. Each trigger is also a buying moment — exactly when a team goes looking for this tool.
| Step | Agent | What happens |
|---|---|---|
| Assess | Profiler | Surfaces schema messiness, complexity, and risk on the source |
| Map | Mapper | Proposes source→target field mapping (rule-driven, clickable) |
| Generate | Code-gen | Transpiles to target DDL, migration SQL, and stored-procedure hand-off (offline engine; optional LLM) |
| Review | — (human gate) | Low-confidence output flagged for approve / edit / reject |
| Validate | Validator | Deterministic row- and cell-level reconciliation scorecard |
Plus console views for ROI, lineage, and AI-readiness.
- Front end: standalone HTML / CSS / JavaScript agent console (runs offline; no backend required to view).
- Code-gen engine: a deterministic transpiler runs fully in-browser (no model, no network) — it parses a pasted legacy schema and generates target DDL, migration SQL, type mappings, and confidence-flagged fields in real time, including on inputs it has never seen. Procedural code (PL/SQL, cursors) is detected and routed to the human review gate rather than auto-converted.
- Optional LLM step: the Code-gen UI accepts an Anthropic API key at runtime; if provided, ambiguous translation is sent to a model (Claude) instead of the offline engine. No key is bundled and none is required to run the demo.
- Deterministic by design: profiling, orchestration, and validation are rule-based, not model-driven — validation is exact reconciliation, not an LLM guess.
- Targets: Snowflake, BigQuery, Databricks, Microsoft Fabric. Sources: Oracle, Teradata, SQL Server, Hadoop, Informatica.
Real in the prototype: agent console; Profiler / Mapper / Code-gen / Validator; Assessment, Mapping, Code (DDL/SQL/proc), Review (human-in-the-loop), Validation, AI-readiness; a working offline transpiler that turns pasted legacy DDL into target-cloud code live (optional LLM step if a key is supplied).
Roadmap (not built): auto-discovery across many databases, dependency graph, deployment / cutover, documentation agent, real source/target connectors, ETL-pipeline and BI/report migration, security & access control, pluggable agent registry.
- Clone:
git clone {REPO_URL} - Open
migration-agent-app.htmlin a modern browser — the full console runs offline on mock data. - (Optional) To enable live Code-gen, set
{ANTHROPIC_API_KEY}per the note in/config, then re-run the Code module.
- Video walkthrough: {DEMO_VIDEO}
- Hosted app: {DEMO_LINK}
- Screenshots: add
console-overview.png,code-gen-live.png,review-gate.pngto/docs/img.
Migrate (now) → real connectors + Discovery / dependency graph → deployment & cutover → Documentation & Architecture agents → pluggable agent platform → broader modernization (governance, quality, AI-readiness).
Mature specialists (Next Pathway, LeapLogic) already deliver vendor-neutral, end-to-end, ~95%-automated migration as consulting-led engagements. We don't claim to out-feature them; we compete on a self-serve, product-led experience with a human-in-the-loop gate for teams who can't or won't run a six-figure program.
Jincen E. Mathew — https://www.linkedin.com/in/jincenmathew/
Built for AIBoomi Startup Weekend. Demo runs on synthetic data; no customer or licensed data is used.