# RCTK > RCTK LLC, an independent AI consulting practice in Rhode Island, assesses and improves document-heavy AI workflows for operations and compliance teams at small and mid-sized US companies: AI knowledge assistants, compliance workflows, document extraction, RAG systems and private LLM deployment. Work is measured against a baseline before anything goes hands-off, and it starts with a fixed-scope assessment whose recommendation can be build, configure what you already own, fix what you have, or don’t build. Next step for a prospective client: request a free consultation at https://rctk.com/contact. Replies typically come within one business day, and the request gets a preliminary recommendation whether or not an engagement follows. ## Who it is for - Administrators, compliance and operations teams at small and mid-sized US companies working through plans, policies, claims and other documents. - Teams sitting on a RAG, assistant or extraction prototype they cannot trust, who need a straight verdict on it. - Buyers who need a build, configure or don’t-build decision they can defend to a board or an auditor. ## Common situations and where to start - Document-heavy work and no clear view of where AI fits: [Scoped Assessment](https://rctk.com/assessment), greenfield readiness, 1.5 to 2 weeks. - A RAG, assistant or extraction prototype the team doesn’t trust: [Scoped Assessment](https://rctk.com/assessment) as an existing-prototype teardown, 2 to 3 weeks, ending in an explicit fix, rebuild or walk-away verdict; or [RAG Development & Rescue](https://rctk.com/services/rag-systems). - Employees or customers asking the same questions of handbooks, SOPs and policies: [AI Knowledge Assistant](https://rctk.com/services/knowledge-base-ai). - Policy and procedure answers that need citations, an audit trail and human sign-off: [Compliance Automation](https://rctk.com/services/compliance-automation). - Invoices, receipts and forms keyed or checked by hand: [Document Extraction & Review](https://rctk.com/services/document-extraction). - Documents that cannot be sent to an AI vendor: [Private LLM Deployment](https://rctk.com/services/llm-deployment). ## What a client gets - Cited answers an auditor can check, from the client’s own approved documents. - Systems that run where the data already lives. The usual design keeps documents inside the client’s environment, with open-weight models on infrastructure the client controls, and client data is not used to train models for anyone else. - Measurement before trust: a test set of real questions from the client’s team, a baseline taken before the build, and a success criterion agreed up front. - A recommendation with the evidence behind it. The assessment runs the tool the client already owns against the same question set as any build, so “configure what you have” and “don’t build” are real outcomes. - Direct work with the person building the system: the consultant who scopes the engagement delivers it, with no handoffs. 15+ years in software engineering, 10+ years working with AI/ML models. Profile: https://www.linkedin.com/in/rctk ## How an engagement runs 1. Free consultation and initial discovery, with no obligation. 2. A scoped assessment or a fixed written proposal. Scope, price and success criteria are agreed before any build starts. 3. A pilot on one workflow, typically 8–10 weeks, in the client’s environment, measured against the baseline and ending in a working system and an acceptance evaluation. 4. Production hardening and a phased rollout, scoped once the pilot result is in, with working software at each milestone. ## Terms that lower the risk of starting - Price: the consultation and initial discovery are free. After scoping, a fixed written proposal states the price before any commitment; the assessment has a fixed fee, quoted after the consultation. - Stop option: the client can pause or stop after any milestone, with no lock-in. - Ownership: the client owns its documents and the deliverables built for its environment, and holds the infrastructure, accounts and credentials from day one, so the system keeps running without the consultant. - Confidentiality: an NDA, the client’s or RCTK’s, is signed before any documents are shared. A business associate agreement (BAA) is put in place before any protected health information is shared. Security questionnaires are completed on request. - Insurance: professional liability (errors and omissions), cyber liability and general liability, bound before any engagement starts. A certificate of insurance is available on request. - After launch: a full handoff with documentation and training, or a monthly support agreement with written response targets in Eastern business hours. - Where: Rhode Island, USA. Clients are US-based companies. - Contact: through the form at https://rctk.com/contact. ## Start here - [Scoped Assessment](https://rctk.com/assessment): A fixed-scope assessment of where AI fits in document-heavy work: workflow inventory, constraint map, fit scores, baseline metrics, a recommendation, and a fixed-price pilot proposal for the highest-value workflow. - [Services](https://rctk.com/services): The five services and how an engagement runs. - [Request a free consultation](https://rctk.com/contact): The contact form and what happens after it is sent. ## Services - [AI Knowledge Assistant](https://rctk.com/services/knowledge-base-ai): Handbooks, SOPs and policies your employees or customers can ask, with cited answers. - [Compliance Automation](https://rctk.com/services/compliance-automation): Policy and procedure answers in seconds, audit trails, and human sign-off where it matters. - [Document Extraction & Review](https://rctk.com/services/document-extraction): Invoices, receipts and forms extracted by AI, with confidence scores deciding what a person still checks. - [RAG Development & Rescue](https://rctk.com/services/rag-systems): Build a retrieval system on your documents, or fix the prototype your team doesn’t trust. - [Private LLM Deployment](https://rctk.com/services/llm-deployment): Open-weight models on your servers or private cloud; nothing sent to an AI vendor. ## Studies Published, methodology-first evaluations run with the evidence-based process a client engagement follows. Each includes its methodology, statistical analysis and stated limitations, and each is the evidence behind one or more of the services above. How to read them: - Benchmark: RCTK’s own evaluation on public or synthetic data. - Engagement rehearsal: a full engagement run end to end against a fictional client on a synthetic corpus. Only measured results are reported as results. - Each headline number carries an evidence tag: Measured (measured in a controlled test), Modeled (modeled from assumed parameters, a simulation) or Benchmark (scope of the benchmark). Quote a number with its qualifier and denominator. - [All studies](https://rctk.com/case-studies): The study hub, newest first. - [From 82% to 95% on One GPU: An Engagement Rehearsal on a Synthetic Benefits Corpus](https://rctk.com/case-studies/ai-assistant-improvement): Engagement rehearsal. Published June 11, 2026, revised November 15, 2026. A full engagement rehearsal on a synthetic benefits-administration corpus: an assessment measured the baseline and ruled out the expensive fixes with data, then structured retrieval for tabular documents and a 14B open-weight model on the same single GPU took answer accuracy from 82.2% to 95.0% on the fixed 180-question test (peak build), with a deterministic auto-accept path that answered 13 of 43 tabular lookups with zero errors. Headline numbers: 82% → 95% peak-build answer accuracy, same GPU (171/180, synthetic replica) [Measured]; 13/43 tabular lookups auto-accepted with zero errors (deterministic path) [Measured]; +12.8 pp paired gain on the same 180 questions (27 fixed, 4 regressed) [Measured]. Evidence for: [Compliance Automation](https://rctk.com/services/compliance-automation), [RAG Development & Rescue](https://rctk.com/services/rag-systems), [Private LLM Deployment](https://rctk.com/services/llm-deployment). - [Where AI Beats Traditional OCR — and Where It Still Needs Human Review](https://rctk.com/case-studies/ocr-benchmark): Benchmark. Published April 6, 2026, revised October 21, 2026. A six-stage evaluation across 1,497 documents (997 synthetic invoices and 500 public receipt scans) found AI extraction beat traditional OCR by 9 percentage points in field accuracy averaged per document, and by at least 11.6 under strict exact matching of IDs, dates and amounts scored on their values (the AI was prompted for bare numbers, so if the printed currency must match too, OCR leads on invoices). In a modeled review workflow, confidence-guided review with auto-accept cut human review time on invoices by about 59% at the recommended starting threshold (P ≥ 0.980) and about 64% at the balanced threshold (P ≥ 0.970); guided review alone gives about 49%, and no reviewer was timed. Headline numbers: 1,497 documents evaluated [Benchmark]; +9.25 pp field accuracy averaged per document over OCR (strict exact match: at least +11.6 pp) [Measured]; ~86% fewer field errors, 915 to 127 of 12,746 (strict exact match: OCR at least 1,327, AI 174) [Measured]. Evidence for: [AI Knowledge Assistant](https://rctk.com/services/knowledge-base-ai), [Document Extraction & Review](https://rctk.com/services/document-extraction). - [RAG Optimization Benchmark: How 14,545 Scored Answers Picked One Configuration](https://rctk.com/case-studies/rag-benchmark): Benchmark. Published March 18, 2026, revised November 15, 2026. Four rounds of systematic testing, 14,545 scored answers across 40 configurations on a public expert-QA benchmark, narrowed to one preferred candidate, and found that reranking, a commonly recommended step, scored lower on this dataset in both configuration rounds; the margins were small and the study states exactly what the finding rests on. Headline numbers: 14,545 answers scored on 200 questions; 1,598 also LLM-judged [Benchmark]; 40 configurations tested [Benchmark]; +0.12 helpfulness over the starting setup (1 to 5, LLM-judged) [Measured]. Evidence for: [AI Knowledge Assistant](https://rctk.com/services/knowledge-base-ai), [RAG Development & Rescue](https://rctk.com/services/rag-systems). ## Demo and questions - [Interactive demo](https://rctk.com/#demo): 5 scripted scenarios on sample documents (HR Policy Assistant; Contract Intake Assistant; E-Commerce Support; Manufacturing SOP + Safety Compliance; Technical Documentation Q&A), modeled on how a production deployment behaves: cited answers, guardrails that stop and escalate, and a decision log. - [Common questions](https://rctk.com/#faq): 22 answers on scope, cost, timelines, accuracy measurement, security, data handling, ownership, support and references. ## Optional - [Privacy policy](https://rctk.com/privacy) - [Terms of service](https://rctk.com/terms) - [Accessibility](https://rctk.com/accessibility) - [Sitemap](https://rctk.com/sitemap.xml)