InsideBoard AI

AI-assisted data-intake platform, enterprise automotive

A deployed pipeline that ingests messy enterprise Excel feeds and delivers clean, proven data, with the AI kept on a tight leash.

Forward Deployed Engineer, InsideBoard AI · May 2026 to present · Enterprise SaaS, automotive data

Decision Room 01

Deployed client product

A file is cleared for delivery, or it stops with an explanation.

The successful execution is recorded below. Inspect the operating contract behind the outcome without exposing client data or implementation secrets.

89
Expected columns in the recorded model: 16 sec
Recorded validation of 226 rows: Python
Deterministic final decision

Recorded run · successful delivery

Recorded workflow25 sec25 sec · silent

89 cols · verified89 columns · deterministic validation

Explore the recorded run

The operator selects the expected file model before any data is processed.

The recorded client flow: choose the expected schema, upload the source workbook, follow each validation gate, then download the normalized file. End-client identity removed.

Interrogate the system

Input contract

A recurring workbook and an explicit expected model.

The operator selects the expected file model before upload. The workbook is fingerprinted and matched to a versioned structural contract before normalization begins.

Privacy boundary

The model sees column structure only. Cell values remain outside the LLM.

Execution path

01Receive
02Identify
03Validate
04Decide
05Deliver

The problem

An enterprise automotive brand sends recurring, high-column-count Excel feeds through InsideBoard. Templates drift silently between deliveries: columns rename, reorder, and disappear. That breaks downstream normalization and forces slow, error-prone manual checking every cycle.

What I built

I built the full intake platform end to end. n8n stays a thin conveyor; a Python service runs an 11-step pipeline that parses each workbook, proves it against a versioned template registry, and normalizes it to delivery-ready CSV. An LLM agent analyzes structural anomalies and surfaces evidence, but a deterministic integrity belt has the final word: the AI advises, Python decides, so a hallucination can never ship wrong data. The model only ever sees column structure, never cell values.

The outcome

Two people were each spending over five hours a week checking these files by hand to confirm the data was formatted correctly before downstream processing, more than ten hours a week between them. That work is gone. The pipeline validates, normalizes, and delivers on its own, and the team is simply notified that a file cleared or was blocked. A file surfaces only when something is genuinely off, with the exact anomaly, the cause, and the agent's reasoning attached, so the only manual step left is chasing a client who sent bad data. Every run is traced and permanently logged. Deployed and running in the delivery environment.

See it working

The n8n intake workflow: a 5-minute schedule triggers an SFTP scan, then branches to a Slack success or failure alert. n8n stays a thin conveyor; the Python pipeline does the real work.

The client portal after a recorded run: 226 rows validated in 16 seconds, with all 89 expected columns present and in order. End-client identity removed.

Why you can trust it

Multi-agent dev harness: tool-scoped review agents (architecture, operational, security, planning) that vote on every change in sequence, plus a knowledge-graph code index agents query instead of grepping raw source
Tiered context system: a controlled-vocabulary glossary, numbered session-handoff docs, and per-agent persistent memory with automated nightly consolidation, so any fresh AI session rebuilds full project state without re-reading the codebase
Append-only decision log (60+ locked ADRs), plus a security gate that vets every third-party agent or skill for prompt injection before it touches disk
Four-stage CI: 550+ hermetic tests with coverage, SonarQube, per-PR preview builds, and an automated AI code-review gate, behind push-to-main and secret-exfiltration guard hooks

StackPythonFastAPIn8nopenpyxlLiteLLM / AnthropicLangfusePostgreSQLKubernetesArgo CDSlack

Have a project like this?

Book a call More work