GetWidget ships AI development services across the full production stack — LLM apps and agents on Claude and OpenAI, RAG and retrieval over your corpus, sub-second voice agents, intelligent document processing, AI integration into Salesforce / NetSuite / Zendesk, AI governance for EU AI Act and NIST AI RMF, and Flutter mobile apps from the team behind the 4,800-star open-source GetWidget UI kit. Audit-first. Eval-tested. Model-agnostic. Operator-led, not partner-deck.
The GetWidget services catalog covers 12 AI service pillars plus 2 Flutter pillars, each shipped under the same operator playbook: a fixed-scope discovery audit produces a ranked workflow list with eval criteria and a token-cost projection; a 4-6 week pilot ships one workflow end-to-end behind a feature flag with a walk-away point at the eval baseline; continuous monthly engagement covers model swaps, eval drift, and on-call rotation. Each cluster delivers a concrete artifact per pilot: an agent pillar ships a planning loop with tool-use schemas, a chatbot or knowledge-base pillar ships a confidence-gated RAG endpoint, a voice pillar ships a sub-second Realtime API line over Twilio, and an IDP pillar ships a vision-extraction service writing structured JSON. The Flutter cluster ships a release-ready build or an embedded developer pod. Every engagement is model-agnostic across Claude Opus 4.7, Claude Sonnet 4.6, GPT-5, and open-source weights.
Most buyers arrive here looking for "an AI developer" the way you'd look for a backend developer in 2015. The shape of the work is different. A modern production AI build is a stack: a model (Claude Sonnet 4.6, GPT-5, gpt-realtime-2, or a self-hosted Llama 4), a retrieval recipe over your corpus (hybrid pgvector + BM25, reranked), a tool layer (function-calling into your existing APIs), an eval harness with a frozen golden set, and an operations posture covering cost, latency, and audit logs. Every service pillar on this page is one slice of that stack — and most engagements wire several together.
We're model-agnostic by posture, not by indecision. We've shipped Anthropic-only stacks, OpenAI-only stacks, and mixed-vendor stacks where Sonnet does the long-context reasoning and Haiku does the cheap routing. The model is the second decision — the first is whether the workflow is even case-study-shaped. That's what the audit answers: which workflow, which model, which retrieval recipe, and crucially, when the answer is "buy off-the-shelf instead of building." We've told buyers to use a SaaS tool more than once. It costs us a deal and saves them six figures.
The audit narrows the twelve AI service pillars below into the one (sometimes two) you should fund first. That ordering matters: shipping a chatbot before you have a clean retrieval layer is how pilots die. Most engagements run audit → pilot on one workflow → continuous team picking up the next item on the roadmap. We don't sell a platform, we don't gate the eval set, and the engineer who ran your audit is the engineer who ships your pilot.
That's the difference between us and a single-vendor shop. A "Claude consultancy" or an "OpenAI agency" has a tool they're optimising you into. We pick per workflow, publish the math, and ship into your repo on your cloud — your model contracts, your eval set, your audit logs. The page below is organised so a buyer can land on the situation they're in, find the service that fits, and verify the work on a real case study before they ever book a call.
Twelve live AI service pillars, organised by buyer mental model: four for building AI products from scratch, four for integrating AI into existing systems, four for choosing the right model and governing the program. Most engagements start with the audit — we rank your highest-impact candidate and tell you which pillar to kick off first.
Generative AI, LLM agents, RAG, and vision pipelines for mobile, web, and backend. Eval-tested, token-optimised, first workflow live in 30 days.
Autonomous agents for customer service, sales enrichment, and back-office ops. ReAct, plan-and-execute, and hierarchical multi-agent recipes shipped in 4–6 weeks.
RAG-grounded chatbots on Claude Sonnet and GPT-5 mini to your web widget, WhatsApp, voice, or Slack. Eval-gated, guardrailed, first chatbot live in 30 days.
Sub-600ms conversational voice agents on OpenAI Realtime, Deepgram, and Twilio. Telephony and mobile-native voice inside Flutter, iOS, and Android.
Connect Claude, GPT-4, and open models to Salesforce, NetSuite, Zendesk, Slack, and the rest of your stack. First integration live in 30 days.
End-to-end AI workflow automation: sales, ops, support, and document workflows live in 6–8 weeks. Cost-of-ownership reported monthly.
Multi-modal IDP pipelines for invoices, contracts, claims, and medical records. Claude Opus 4.7 or GPT-5 vision, HITL queues, confidence bands, ERP integration.
Internal RAG over your docs, tickets, transcripts, and Slack history. Citation-accurate, role-aware, audit-logged. Agent-assist and search built on the same retrieval layer.
Fixed-scope discovery audit, written roadmap on Day 5. Named workflows, walk-away conditions, and a model recommendation grounded in your eval set.
EU AI Act, NIST AI RMF, ISO 42001 — we ship the eval suites, audit logs, red-team findings, and remediation PRs an auditor can actually sign off on.
Anthropic Claude specialists: long-context agents, tool use, Computer Use, RAG, and Claude Code engineering. Daily operators, not slide-deck consultants.
GPT engineering, Realtime voice agents, function-calling workflows, Assistants API, and Codex. Model-agnostic with transparent token-cost math.
The audit ranks your highest-ROI AI candidate, scopes a pilot, and tells you when not to build. Fixed-fee, 1–2 weeks.
Book the auditThe same engineers who ship your AI workflows also build and staff Flutter apps — native-quality cross-platform, AI-augmented, backed by the open-source GetWidget UI library.
Flutter mobile apps, AI inside Flutter, and honest cross-platform engineering from the team that publishes the GetWidget Flutter UI Kit (4,811★ · 23K monthly pub.dev downloads).
Hire dedicated Flutter engineers from the team behind the GetWidget library — AI-augmented delivery with Claude Code in our repos. Audit-first, no marketplace markup.
Every metric on this page was drawn from shadow-mode logs, frozen eval sets, or 30+ day A/B tests on shipped engagements. The brand names are changed at client request; the math is not.
Every engagement starts with a fixed-fee audit. One to two weeks, fixed-fee, written deliverable. We shadow the actual work (not a demo, not a deck), score your candidate workflows on ROI, risk, and time-to-ship, and hand back a written roadmap on Day 5 with named deliverables, a model recommendation grounded in your eval set, and the walk-away condition that would end the pilot. Roughly a third of audits end with "don't build this — buy this off-the-shelf tool." We charge for the audit either way; we don't charge for the pilot if the audit says it isn't shaped to ship.
If the audit clears, the pilot is fixed-bid: fixed-bid, 4–8 weeks, one workflow end-to-end. We build the eval harness before we build the prompt, because every production AI system is a regression-tested artifact and the eval set is the only thing that catches model drift, retrieval drift, and prompt-rot before your users do. The pilot ships behind a feature flag with a logging spine, a fallback runbook, and a published kill-point: if the metric on the eval set doesn't move by week six, you don't pay for Phase 2. The kill-point is signed off in writing before week one.
Model selection happens per workflow, not per engagement. We've shipped Claude-only stacks, OpenAI-only stacks, and mixed stacks where Sonnet does the reasoning and Haiku does the cheap routing. The model is picked on the eval set after the retrieval recipe exists; never before. When a new model lands (and they land every six weeks now), we re-run the bake-off — the eval set is the constant, the model is the variable. You own both at the end of the pilot.
We don't quote everything as a six-month engagement. Most clients start with a fixed-fee audit, ship one workflow on a pilot, then move to monthly for the roadmap. Pick the entry point that matches your certainty level.
Map the highest-ROI workflow or build candidate before you commit a budget.
One workflow or integration, end-to-end, with eval data — not a demo.
Embedded squad shipping the next item on your roadmap each sprint.
30 minutes — we review your highest-ROI workflow or build candidate, recommend a model + retrieval recipe, project token and run-cost, and tell you honestly whether it's worth building or whether an off-the-shelf platform covers it. No deck, no obligation to proceed.
Each service pillar feeds back into an industry, a case study, or the open-source kit. Start anywhere.
Single canonical entry for our AI work — pillars, methodology, benchmarks, open-source harness.
Dated, reproducible benchmarks on RAG retrieval and agent reliability. Cost on the same axis as quality.
How we ship AI: audit → pilot → continuous, gated on weekly evals. Standards-mapped (NIST AI RMF, ISO 42001, EU AI Act).
By vertical: healthcare, legal, fintech, ecommerce, education, manufacturing.
Six published engagements with eval data, latency budgets, and cost math.
4,800-star Flutter UI kit. Heritage OSS behind the AI engineering team.
MIT eval harness (Ragas + promptfoo + agent rubrics). Same code that produces the public benchmarks.