Process

How enterprise AI gets deployed.

Building enterprise AI is not about dropping a chatbot into an iframe. It requires a rigorous, phased methodology designed for traceable answers, tight security and measurable results.

How an engagement runs

Phase 1 · Discovery

Establish what is buildable

We do not write production code in week one. We establish what your data and systems can actually support, and what a correct output looks like.
- —Process mapping with the people who do the work today
- —Constraint capture: data residency, approved providers, audit requirements
- —Data readiness — the state of your APIs, databases and documents
- —An evaluation set and an agreed threshold, written down
Phase 2 · Build

Build against real systems

Retrieval, agent loop and integrations built against your actual systems rather than a sandbox, with the evaluation harness running from the first week.
- —Retrieval or agent architecture on real data
- —Scoped, audited tool access — read-only by default
- —Guardrails: untrusted content stays data, never instructions
- —Measured iteration against the evaluation set
Phase 3 · Production

Ship it, then hand it over

Shadow mode first, so you can compare the system against current practice before it acts. Then monitoring, regression gates and a documented handover.
- —Shadow run against live traffic, measured not assumed
- —Regression gates in CI so quality cannot drift unnoticed
- —Cost and quality monitoring with alerting
- —Runbooks and handover — your team can own it

What the work looks like

agent/definition.py PYTHON

agent = Agent(
                    model=MODEL_PINNED,     class="text-dim"># upgraded deliberately, behind an eval gate
                    tools=[
                        search_docs,        class="text-dim"># read-only
                        fetch_account,      class="text-dim"># read-only, scoped to the caller
                        create_ticket,      class="text-dim"># write, rate-limited
                    ],
                    class="text-dim"># Retrieved content is data, never instructions.
                    untrusted_sources=[DOCS, TICKETS],
                    max_steps=class="text-plasma">8,            class="text-dim"># bounded - no runaway loops
                )

Agents are defined as code with explicit, scoped tools. Read-only by default; anything destructive is either absent or gated behind confirmation.

ci/evals.yml YAML

- name: Evaluation suite
                  run: |
                    uv run evals/run.py \
                      --set golden \
                      --baseline main \
                      --fail-under 0.92

                # Retrieval is measured separately. A generation score that
                # drops because retrieval regressed is a misleading signal.
                - name: Retrieval recall
                  run: uv run evals/run.py --set retrieval --fail-under 0.88

Quality is a merge gate, not an opinion. A pull request that drops the score on the golden set does not land. Thresholds are illustrative; yours are agreed in discovery.

CUTOVER PROTOCOL

From shadow mode to full autonomy

We push the architecture to production, initially in a Human-in-the-Loop (HITL) shadow mode. Once the agent meets the agreed accuracy threshold on live traffic, we widen its autonomy step by step, with monitoring and a human escalation path in place throughout.