CAPABILITY / 08

Applied AI & ML

In production

AI that touches real systems through a scoped identity, with traces and an audit trail around it.

Updated 15 Nov 2026

01 / PROOF

The measured part

11
read-only tools in the Job Portal AI connector
Source
92
read-only tools in the Scheduler ops/HR connector
Source
AES-256-GCM
CHECK-enforced encryption in the AI trace store
Source

02 / WHAT I DO

In practice

Build read-only AI access to production data with per-call identity checks.
Record prompts and completions with AES-256-GCM encryption and per-worker erasure.
Keep connector calls behind a scoped identity and an audit trail.
Train engineers on MLOps: pipelines, containers, deployment and monitoring.

03 / WHERE I'VE DONE IT

The systems behind it

RVM CAD Job PortalRead-only MCP connector with 11 tools and 141 tests.RVM CAD Job PortalPer-call identity re-check on every MCP request.RVM CAD Job PortalOAuth 2.1 + PKCE and rotating hashed refresh tokens.RVM CAD SchedulerOps/HR connector with 92 read-only tools and an append-only audit trail.BadaBhaiEncrypted AI trace store with per-worker erasure.FUTURENSETaught MLOps to working engineers and postgraduate cohorts.

04 / STACK

What it runs on

Python
FastAPI backend, ingestion and analytics
See the system
Gemini
Primary model, with a Claude fallback
See the system
Claude
Fallback model
See the system
MCP
Read-only connector with 11 tools
See the system
Azure ▸Azure AI Foundry (in my roles)In my roles
Azure AI Foundry in my roles (self-reported)
Airflow (in my roles)In my roles
Data pipelines and scheduling in my roles

05 / LIMITS

What this does not claim

Model training at Boven is not detailed here. No accuracy figure is claimed.

06 / RELATED CASE STUDIES

All work

Built and testedRVM CAD Job Portal and RVM CAD Scheduler

Both connectors are merged with full test suites. Production activation and the CEO acceptance test had not been verified when this was written (20 Oct 2026).

Doors, not keys: read-only AI access to production data

Two connectors that let an AI assistant answer operational questions from live systems — without a database credential, without write access, and without ever returning a half-right number.

11 read-only tools on the job portal connector

ProductionBadaBhai

Shipped in August 2026; the migration was applied to production with the release. Six AI call sites write to it.

An AI trace store that treats its own data as dirty

Recording every prompt and completion on an AI hiring platform, so bad answers can be investigated — encrypted at rest, erasable per worker, and built so raw text cannot reach the recorder by accident.

6 AI call sites traced

07 / QUESTIONS

Short answers

How does KP keep an AI assistant away from raw production data?

The connector is read-only and scoped: OAuth 2.1 with PKCE, hashed rotating refresh tokens, a per-call identity re-check and an audit trail. The Job Portal version has 11 tools; the Scheduler ops/HR version has 92.

What does the AI trace store protect?

Prompts and completions are encrypted with AES-256-GCM under a CHECK-enforced ciphertext shape and are erasable per worker.

Need this on a real system?

Hiring for a role, or scoping a short engagement? The Hire page has the role families and contact; Consulting has the scoped offers.

For hiring managers Consulting