The engineering under every system

When we reach for it

Three real triggers: proprietary data with no off-the-shelf equivalent, hard latency or edge constraints, or IP ownership requirements. We check for all three before recommending custom model work, because at least two are usually absent.

When we reach for it

Tone of voice, structured-output reliability, narrow task specialization, and cost reduction where a small fine-tuned model replaces a big general one. Distillation is the quiet win: a frontier model generates the training data, a small model serves it at a fraction of the unit cost.

When we reach for it

Self-hosting wins in exactly three cases: data sovereignty, very high sustained volume, or a fine-tuned small model. Then we reach for vLLM, quantization, and careful prompt-cache and KV-cache engineering, and we benchmark the unit economics before and after.

When we reach for it

Internal knowledge assistants, support copilots, contract and policy Q&A, research tools. We start with pgvector inside the Postgres you already run, use hybrid search because pure vector retrieval misses exact terms, and rerank because it is the cheapest quality win in most systems. Retrieval quality is measured separately from generation, and before it.

When we reach for it

Back-office automation, research and report pipelines, customer-ops copilots. Every agent we ship runs on the same architecture: plain-text directives your team can edit, model judgment spent only at decision points, deterministic scripts for everything else, an approval gate before anything irreversible, and an audit log of every action.

When we reach for it

Every engagement. A golden dataset before we build, regression gates before we ship, tracing in production after. The client keeps the eval suite; it is part of the deliverable, because a system you cannot measure is a system you cannot trust.

When we reach for it

Tabular prediction: churn, lead scoring, risk, pricing, demand forecasting. Gradient-boosted trees still beat deep learning on most business data, run for pennies, and produce feature attributions an auditor can actually read.

When we reach for it

Document-heavy workflows are the most common win we see: invoices, contracts and forms into structured, validated data, with a human review queue on the output. Vision-language extraction with schema validation now beats bespoke OCR at most volumes, and Whisper-class transcription feeds call analytics and voice agents.

When we reach for it

Every system we ship lands on infrastructure we can defend: serverless Python on Modal, dashboards on Vercel, Postgres on Supabase, CI with eval gates on GitHub Actions. Production hardening of a stalled internal pilot is a standalone engagement we take often.

When we reach for it

Every build: row-level security in the database, per-client isolation, model APIs under no-training terms, approval gates and audit logs on anything that acts. Customer-facing deployments add prompt-injection testing and output filtering before launch, not after the first incident.