LLM Integration
Claude, GPT, and open models. Tool use, structured output, streaming, function calling, prompt caching, model routing, and the boring plumbing that makes them production-safe.
Newton Technology · Est. MMVI
Production AI, engineered into
software that ships.
Newton Technology is the practice of Joe Newton — twenty years building for the web, now focused on AI-native products: integration, retrieval, agents, evals, and the unglamorous engineering that turns a clever demo into a reliable system.
§01 — Principles
Three things I refuse to compromise on.
AI work in 2026 is overflowing with confidence and underflowing with measurement. The practice below is what separates a feature that ships from a demo that lingers.
-
01
Engineering, not theater.
A weekend prototype is the easy 20%. The hard 80% is latency budgets, retries, idempotency, observability, cost ceilings, and graceful failure on the worst inputs your users will ever produce. That’s the work I do.
-
02
Evals before features.
If we can’t measure it, we can’t ship it — and we certainly can’t improve it. Every AI surface I build starts with an eval harness, a regression set, and a number that has to go up before anything goes out.
-
03
Latency is design.
Streaming, caching, optimistic UI, and graceful degradation aren’t back-of-the-sprint optimizations — they’re the product. Users feel every millisecond, and a fast wrong-ish answer often beats a slow perfect one.
§02 — Capabilities
What I’m hired to build.
A working list, not a brochure. If your problem is shaped roughly like one of these, we should talk.
Retrieval & RAG
Vector search, hybrid retrieval, document pipelines, chunking strategies that don’t butcher your data, and citations users can actually trust.
Agents & Workflows
Multi-step reasoning, tool orchestration, human-in-the-loop checkpoints, and the guardrails that keep an agent from running up a four-figure bill on a Tuesday.
Evals & Observability
Eval harnesses, regression suites, golden datasets, traces, and cost dashboards. A feature without an eval is a wish; a feature with an eval is a system.
Web Applications
Full-stack work on modern web stacks: TypeScript, React, Next.js, Postgres, edge runtimes. I’ve been doing this part since long before it was fashionable.
Migrations & Integrations
Bringing AI into an existing product without ripping it apart. Surgical changes, careful rollouts, real measurement — not a rewrite dressed up as a feature.
§03 — Method
How an engagement actually runs.
Five steps. No deck-ware, no “discovery phases” that last a quarter, no surprises in week ten.
-
01
Scope
Understand the problem, not the prompt. We agree on what success looks like, what we’re measuring, and what we’re explicitly not doing yet.
-
02
Prototype
Working code by the end of week one. Always. Real models, real data, real interface — even if it’s ugly and held together with TypeScript.
-
03
Eval
Build the measurement before the polish. The eval becomes the spec; the spec becomes the regression suite; the regression suite is what lets us move fast later without breaking things.
-
04
Ship
Production from day one, behind a flag if needed. No six-month staging environments. Users find things evals never will.
-
05
Iterate
Real traces, real failures, real wins. We tune prompts, swap models, add tools, prune scope. The eval number goes up. We do it again next week.
“If I have seen further, it is by standing on the shoulders of giants.” — Isaac Newton, 1675
§∞ — Note from the practice
I started programming on a beige tower in elementary school and never quite stopped. Twenty years professionally now — the web before it had standards, JavaScript before it had types, AI before it could write a paragraph on its own.
Newton Technology is the name I give to the work I most want to do: careful, measured engineering applied to a field that has more hype than rigor. If that sounds like the kind of partner you’re looking for, the contact section is right below.
§04 — Contact