Full-stack AI, not just a model

Retrieval is where the majority of enterprise AI value sits, and where the majority of prototypes quietly fail. The gap is almost never the embedding model. It is ingestion: tables flattened into unusable prose, multi-column PDFs read in the wrong order, OCR failing silently on scanned pages and producing plausible garbage that gets indexed and cited.

We build ingestion as the first-class component it is. Layout-aware parsing preserves table structure and reading order. OCR runs behind a confidence gate that routes low-confidence pages to human review rather than degrading the index silently. Chunking follows document structure — a clause, a table, a policy section — with contextual headers prepended so each chunk is interpretable on its own.

Retrieval itself is hybrid by default. Dense vectors catch paraphrase; BM25 catches the exact entity names, account numbers and defined terms that dense retrieval reliably confuses. Fused and cross-encoder reranked, then generated under enforced citation: every claim resolves to a document, page and region a human can verify in one click, or the response is rejected by the validator and escalated. Abstention is a first-class, calibrated output — a system that cannot say "not found in this file" is not deployable in a regulated workflow.

Forecasting in real operations is rarely a time-series textbook problem. It is a hierarchy problem, an intermittency problem and a promotions problem simultaneously — and the three interact. A model that nails chain-level demand and cannot be reconciled to store level is unusable, because finance plans against one number and execution happens against the other.

So we forecast distributions, not points, and reconcile by construction. Quantile objectives give replenishment policy the service-level quantile it actually needs — essential for the long intermittent tail where a mean forecast is actively misleading. MinT reconciliation makes every level of the hierarchy sum, which ends a surprising amount of organisational argument. Promotions are explicit features owned by a calendar, never noise absorbed into a trailing average.

Gradient-boosted trees remain the right default for tabular forecasting at this scale — they outperform deep sequence models on most real retail and operational data, train in minutes rather than hours, and expose feature contributions that a planner can interrogate. That last property is what drives adoption: a forecast that shows its top drivers and last year's actual gets trusted; an unexplained number gets overridden into irrelevance.

This is the half of the work that determines whether a model is still running in eighteen months. It is also the half most consultancies leave for the client to figure out, which is how organisations end up with a Jupyter notebook in production and one person who knows how to restart it.

We ship reproducible pipelines: versioned data, versioned code, versioned models, versioned prompts, with a registry that can tell you exactly what was serving on any given date. Evaluation runs in CI as a gate — a change that regresses a tracked metric beyond tolerance does not merge. That single control prevents more production incidents than any amount of monitoring, because it catches the regression before it ships.

Deployment topology follows your constraints rather than our preferences. Roughly half our deployments run inside a customer VPC; we regularly deploy fully on-premises, including air-gapped facilities with no outbound network access. That constraint is identified in week two of the engagement, because it determines model selection — and discovering it in week eight is how AI projects die.

Then we hand it over properly: runbooks, dashboards, alert thresholds, an on-call escalation path, and training for the team that owns it. The measure of a good engagement is that you do not need us afterwards.