# COXIT

> Engineer-led AI/ML development studio building custom AI for document- and drawing-heavy workflows — computer vision, LLMs, and AI agents. Production systems shipping for organizations such as Internet Archive and Stevens Industries. Based in the USA (Casper, Wyoming) and Ukraine (Lviv).

COXIT designs and ships custom AI systems end to end: scoping and proof of concept, model and pipeline development, and long-term production support. Independent AI audits are also offered. Engagements are led by senior engineers, and the team also provides staff augmentation (team extension) for AI/ML, Python, and C++ work.

## Services
- [AI Audit](https://coxit.co/service/ai-audit/): Independent technical review of an existing AI system or build plan — accuracy numbers, defect list, and a prioritized fix plan with time estimates.
- [AI/ML Software Development](https://coxit.co/service/machine-learning/): Custom AI systems — computer vision, LLMs, and agents for document-heavy workflows.
- [Team Extension](https://coxit.co/service/team-extension/): Senior engineers embedded on your team in 2–4 weeks at an all-inclusive rate.
- [LLM Evaluation](https://coxit.co/llm-evaluation/): Evaluation systems that measure how accurate an LLM system is — ground truth datasets, metrics, and automated runs that show whether each model or prompt change improves it.

## Case studies
- [Production Takeoff Tool for Stevens Industries](https://coxit.co/cases/estimation-stevens-part3/): Custom AI that cut cabinet estimating from ~6 hours to ~10 minutes at a large U.S. casework factory.
- [Spec Book AI Agent for Stevens Industries](https://coxit.co/cases/ai-powered-document-agent/): AI agent that reads architectural spec books to extract requirements from drawing-heavy documents.
- [Stevens Industries POC](https://coxit.co/cases/discovery-stevens-part1/): Discovery phase for machine-learning-powered estimation software in manufacturing.
- [Stevens Industries POC, Part 2](https://coxit.co/cases/poc-stevens-part2/): Proof of concept validating the estimation approach before custom development.
- [Multi-Agent Investment Research Platform for BRDGE Insights](https://coxit.co/cases/strem/): Multi-agent LLM system for fintech investment research.
- [AI-Powered Investment Research Platform — Foundation](https://coxit.co/cases/strem-discovery/): Discovery and architecture for the investment research platform.
- [Internet Archive](https://coxit.co/cases/archive-org/): Python engineering supporting large-scale book digitization for the world's largest online library.
- [Xperi](https://coxit.co/cases/xperi/): Multi-year embedded and backend engineering team extension.
- [Tempest](https://coxit.co/cases/tempest/): Backend engineers improving data parsers for a privacy-focused browser.

## Industries
- [Manufacturing](https://coxit.co/industry/manufacturing/): AI for estimating, takeoffs, submittals, and document workflows in woodworking and casework manufacturing.

## Resources
- [CaseVBench](https://coxit.co/ai-drawing-benchmark/): How accurately current AI models interpret architectural drawings and identify cabinets, countertops, floor plans, elevations, and callouts, scored against expert markup of the same sheets.
- [AI Adoption in Woodworking 2025 Report](https://coxit.co/report/): Insights from 40 AWI members on AI adoption, operational challenges, and where automation delivers the most value.
- [Blog](https://coxit.co/blog/): Practical articles on LLMs, prompting, RAG, machine learning, and AI for manufacturing.
- [Careers](https://coxit.co/careers/): Open engineering roles and how the team works.
- [Full site text](https://coxit.co/llms-full.txt): Every article in full, plus case study and service detail, as a single plain text document.

## Blog articles

Written by the engineers who built the systems described. Grouped by subject, newest first.

### AI for Drawings
- [GPT-6 Astra Scores 92% on Architectural Drawings: What Does It Mean for AI Takeoffs?](https://coxit.co/blog/gpt-6-astra-architectural-drawings-benchmark/): We ran eleven multimodal AI models over 119 architectural drawing pages and 1,430 hand-labelled objects. GPT-6 Astra scored 92% F1, up from 62% in August.
- [How Accurate Is AI at Reading Architectural Drawings? Benchmark Results for 6 AI Models. Part 2](https://coxit.co/blog/ai-architectural-drawings-benchmark-part-2/): We ran six multimodal AI models over 119 architectural drawing pages and 1,430 hand-labelled objects. Qwen3.8-Max led at 62% F1 — but what you ask the…
- [How Accurate Is AI at Reading Architectural Drawings? A Benchmark of 6 AI Models. Part 1](https://coxit.co/blog/ai-architectural-drawings-benchmark-part-1/): Multimodal models can find many casework objects on architectural drawings, but accuracy swings with object type, drawing resolution, and workflow — and…
- [6 Questions About AI Takeoffs for Casework Shops](https://coxit.co/blog/6-questions-about-ai-takeoffs-for-casework-shops/): You're about to spend significant money and management attention on an AI system. Whether it succeeds or becomes expensive technical debt depends almost…
- [From 6 Hours to 10 Minutes: How AI Transformed Estimating at the Largest U.S. Casework Factory](https://coxit.co/blog/ai-cabinet-estimating-takeoff/): One of largest U.S. manufacturer of commercial cabinetry operates a highly automated factory: CNC machines, automated material handling, and robotics. One…

### ML Engineering
- [Evaluating RAG using LLM-as-a-Judge Independent Evaluation, Automatic Insights Collection](https://coxit.co/blog/evaluating-rag-using-llm-as-a-judge-independent-evaluation-automatic-insights-collection/): The entire approach is open-sourced and is accessible as a GitHub repository. Evaluation? Why RAG output still needs checking In my latest article…
- [Chatting with Large PDFs (100–500 Pages): Using RAG with OpenAI Embeddings (Local vs. API)](https://coxit.co/blog/chatting-with-large-pdfs-100-500-pages-using-rag-with-openai-embeddings-local-vs-api/): At COXIT, in 2025, we observe a trend when more and more potential customers ask us to implement a way to chat with massive documents or even sets of…
- [How to set up monitoring of CPU and memory usage for C++ multithreaded application with Prometheus, Grafana, and Process Exporter](https://coxit.co/blog/how-to-set-up-monitoring-of-cpu-and-memory-usage-for-c-multithreaded-application-with-prometheus-grafana-and-process-exporter/): Have you ever wondered how to know how much CPU and memory resources is your app taking during a week? Or how to collect and visualize this data to get…
- [How To Integrate Valgrind into GitHub Actions?](https://coxit.co/blog/how-to-integrate-valgrind-into-github-actions/): Have you ever got into the situation that after you fixed memory leaks with Valgrind you see Definitely lost again after a while? Then this CI setup may…
- [Differences in Prompting Techniques: Claude vs. GPT](https://coxit.co/blog/differences-in-prompting-techniques-claude-vs-gpt/): Introduction When working with language models (LLMs) like Claude and GPT, the effectiveness of prompts can vary significantly based on the model's…
- [Experiments with different LMMs](https://coxit.co/blog/experiments-with-different-lmms/): MythoMax 13B (nitro) Code: gryphe/mythomax-l2-13b:nitro Description: One of the highest performing and most popular fine-tunes of Llama 2 13B, with rich…
- [Minimizing Randomness in GPT Responses: A Guide to Using Seeds, Top_p, and Monitoring System Fingerprints](https://coxit.co/blog/minimizing-randomness-in-gpt-responses-a-guide-to-using-seeds-top_p-and-monitoring-system-fingerprints/): When working with large language models (LLMs) like GPT, a major challenge is ensuring consistency in responses. These models are inherently stochastic…
- [Prompt Engineering through Structured Instructions and Advanced Techniques](https://coxit.co/blog/prompt-engineering-through-structured-instructions-and-advanced-techniques/): Introduction Language models (LLMs) are powerful tools for a variety of tasks, but their effectiveness is highly dependent on the design of prompts. This…

### COXIT Newsletter
- [July 2026: Woodworking Innovations in Italy, a 40 Under 40 Award, and a New Client](https://coxit.co/blog/woodworking-innovations-italy-40-under-40-major-new-client/): Notes from XYLEXPO in Milan on how much shop-floor business software still lives on legacy desktops, a Wood Industry 40 Under 40 nod, and a PoC quoting…
- [COXIT Co-Founder Volodymyr Hresko Named to Woodworking Network's 2026 Wood Industry 40 Under 40](https://coxit.co/blog/coxit-co-founder-volodymyr-hresko-named-to-woodworking-networks-2026-wood-industry-40-under-40/): The custom AI developer agency is recognized for bringing computer vision and AI agents to document- and drawing-heavy manufacturing workflows in the wood…
- [June 2026: How We Built a Multi-Agent Investment Research Platform](https://coxit.co/blog/how-we-built-multi-agent-investment-research-platform-strem/): Two months of audit before a line of product code, then a team of specialised agents under an orchestrator. Strem now monitors 3,000+ US securities around…
- [March 2026: AI Audit Launching, Swiss Market Research, Canaries Workation](https://coxit.co/blog/2026-ai-audit-swiss-market-research-canaries-workation/): A look back at 2025 — three anchor clients, 40 woodworking shops surveyed, four Swiss projects — and the new AI Audit service we are launching for teams…
- [October 2025: AI Engineers Who Don't Accept Offers, Document Extraction, and LLMs vs SLMs](https://coxit.co/blog/ai-engineers-who-dont-accept-offers-document-extraction-llms-vs-slms/): Over 200 candidates sourced, two offers made, zero accepted — what we changed in our hiring process, the document-extraction request we keep hearing, and…
- [March 2025: How COXIT Transformed Estimation at Stevens Industries](https://coxit.co/blog/how-coxit-transformed-estimation-at-stevens-industries/): Stevens Industries' Continuous Improvement Manager on what changed when AI started reading their architectural drawings, and why pre-estimating that took…

## Contact
- Email: volodymyr@coxit.co
- [Sitemap](https://coxit.co/sitemap_index.xml)