DSEL’s AI Automation Lead turns the firm’s own record — standards, procedures, reports, and historical project information — into internal AI assistants engineers can rely on, agentic workflows that take the repetitive, knowledge-intensive work off their plates, and practical methods for testing and evaluating AI outputs before anyone builds on them. Below is that record, already running: the internal-assistant system in production on two sites, the evaluation protocol run in public, the agent infrastructure, and the products that crossed from prototype to production.
Read closely, this is a build-and-govern brief: one person who finds the opportunities with the engineering teams, prototypes fast, hardens what survives into production tools the whole firm adopts, and writes the standards that keep the AI honest. The scarce skill is not any single technology on the list — it is judgment about what to build, grounded in systems that already run. Every claim on this page links to one.
An assistant that answers from the firm’s standards, procedures and project history is infrastructure. One that improvises is a liability. Grounding rules, source filters and retrieval discipline come first.
Small tools that solve a named problem for named people, tested against real work — then the ones that earn it get the production treatment: pipelines, gates, documentation, adoption.
Engineering work gets signed. AI outputs that feed it need test sets, verdicts and a corrections path — methods established before scale, not after the first bad answer.
A retrieval-grounded assistant that leverages an organization’s own knowledge base — documents, standards, procedures, project record — running in production on nikdesign.ca and teambotics.app over a shared backend. The corpus sits in Supabase/Postgres behind an HNSW vector index; a seeding pipeline with deterministic fingerprints keeps it current and auditable (unchanged documents skipped, orphans cleaned, ingestion runs logged); grounding rules are explicit; moderation and rate limiting are design decisions; and a production smoke test exercises the live endpoint against known failure prompts after every reseed. This is the posting’s centrepiece — internal assistants over the firm’s own knowledge — as a running system.
A two-validator verification protocol for AI-generated claims: falsifiable claim atoms, evidence ledgers, independence-grouped validation — including a validator slot deliberately left empty, because a system cannot validate its own output — plus right-of-reply and corrections standards. The Ground News audit (16 claims, two independent validators, cleared with exceptions) is published in full; the product page carries the method and the standards.
A scoring rubric built as explicit acceptance criteria, standardized test data applied identically across every target, and 247 scored exchanges executed against 32 live production deployments of enterprise AI software. Verdicts — 3 PASS / 42 PARTIAL / 202 FAIL — published per deployment with transcripts and evidence handling, so every finding can be checked. Testing, evaluating and monitoring AI-generated outputs, as an executed campaign rather than a framework on a slide.
Source-aligned knowledge maintenance for humans, teams and AI agents: DyKnow watches approved sources — code, docs, websites, tickets, specs — detects when a change stales a knowledge page, drafts the update, and routes it through human review before anything commits, with an audit log. Ships as a CLI, an MCP server (10 tools) and a VS Code extension. An internal assistant is only as good as the record it answers from; this is the layer that keeps the record honest.
CodeBuddy, published on the VS Code Marketplace, coordinates multiple AI coding agents on a shared codebase — file locking, structured handoffs, live session state. BOSS puts a human approval gate between an IDE agent and repository writes. 4Age, my protocol for human-approved agent communication, governs a live agent email mesh — scoped approvals, receipts for consequential acts. Agents that reason through a process, access information and use software tools, under human authority: that is the daily operating model, and the tooling is public.
LTB Buddy turned a document-heavy legal process into a live guided-intake product with a professional clinic channel. Geek Doctor is a Windows-first desktop application (v0.2.0) with a consent architecture for users who never read manuals. FileBuddy is local-first file intelligence — read-only by design, shipped at 0.4.0. Poko ships versioned cross-platform releases. Different domains, one discipline: find the real problem, prototype against it, harden what survives, and put it in people’s hands.
An assistant that answers from your own standards is infrastructure. One that improvises is a liability.
Map the knowledge estate — standards, procedures, report and project archives — and interview the engineering teams for the repetitive, knowledge-intensive tasks they would hand off first. Publish the opportunity backlog, prioritized with the Head of Automation.
The first internal assistant over the standards and procedures corpus, in production for a pilot group — with evaluation sets drawn from real project questions and the verdicts reviewed with the engineers who asked them.
The firm’s AI technology standards, development practices and deployment process, written from the first deployments. The next two automations in build. Guidance sessions for developers and engineering staff underway.