We build LLM and machine learning features into the products and back-office systems you already run. Assistants that write working code, pipelines that process thousands of hours of audio, and automation inside your CRM. We build the checks that catch bad output before users see it.
Six kinds of AI work we have delivered. Each card links to the project where it runs today.
A chat panel that knows your platform, answers how-to questions, writes code or queries for the user and hands off to a person when it should.
Batch jobs that turn long recordings or files into transcripts, subtitles and summaries, sized for thousands of hours of input.
Assistants that reply to customers, book appointments and send follow-ups at any hour, connected to the CRM and ERP you already run.
Classic machine learning where it fits better than a chat model: scoring feedback, watching transactions and forecasting from your own data.
Text and image generation placed in the fields editors already use, so they draft, illustrate and translate without copying between tools.
Help for students and teachers: hints on hard exercises, and tools that build interactive exercises from existing material.
A model that is right most of the time is not a feature yet. The engineering around it decides whether people can trust it. We pick the checks by what a mistake would cost.
Generated code runs in the actual runtime before the user sees it. If it fails, it goes back to the model for another pass. Users get code that has already executed.
Output lands in a review queue. A teacher, editor or operator can edit it, regenerate it, delete it or approve it. Nothing goes out without that step.
The model fills a defined schema with timestamps, labels or fields. Post-processing then checks and fixes formatting, language rules and references.
Short, simple jobs go to a smaller, cheaper model. Long or hard ones go to a stronger model. Quality stays where it needs to be and spend drops.
Lokad’s clients describe forecasting and inventory logic in Envision, Lokad’s own scripting language. New users needed help to write their first scripts. In 2025, after working with us on their CRM, Lokad asked OSKI to put an assistant directly into the platform.
“This is a team of engineers who obviously like what they are doing.”
An educational archive records lectures where speakers switch between Hebrew, Aramaic and English mid-sentence. We built a Kubernetes pipeline that splits long recordings at natural pauses, transcribes each chunk with Gemini on Vertex AI, and stitches the timeline back together.
Each chunk goes to Gemini Pro or Flash depending on its length, with Flash-Lite also available. Language-aware cleanup adds diacritics and checks references to religious sources. A separate module builds HLS video, so every file leaves as a transcript, subtitles, a summary and a stream.
For EducatedOn, an AI learning platform, we added OpenAI text generation to three Umbraco field types and connected the Stability API for images. Generated images are uploaded and placed into content automatically, and everything runs on Umbraco Cloud.
We also publish this kind of work as Umbraco packages on NuGet, so other teams can use it too.
We are not tied to one AI vendor. We pick the model for the task, its cost and where your data needs to live. These are the models and tools behind the projects on this page.
The surrounding software is ordinary engineering: APIs, queues, storage and deployment. We build that in Python, .NET or Node.js, depending on what your system already runs.
AI features touch customer messages, internal documents and business records. We set them up so you stay in control of where that data goes and who can reach it.
Most AI projects stall between a promising demo and daily use. We plan for daily use from the first week: real data, an agreed metric and checks sized to the risk.
Our engineers use AI tools in their own delivery work. By our estimate that makes delivery about 3x faster. This is our own figure, not a third-party one.
We look at where people lose time today: the questions they repeat, the files they process by hand, the replies they type. One task with a clear before-and-after comes first.
A small prototype runs on real examples from your system, not on a demo set. You see where it is right, where it is wrong and what it costs per call.
Runtime tests, schemas, human approval or model routing, chosen for the risk of the task. This is usually where most of the engineering goes.
The feature goes where people already work: your product, CRM, CMS or back office. No separate tool for users to remember.
We track the numbers agreed at the start, such as first-try success, response time or cost per job, and adjust prompts, models and checks.
Assistants, pipelines, automation and content tools across supply chain software, education, e-learning and operations.
An assistant inside Lokad's platform that answers questions, generates Envision code and tests it against the real Envision runtime.
A Kubernetes pipeline that turns hours of mixed-language lectures into timed transcripts, subtitles, summaries and HLS video.
AI services connected to existing CRM, ERP and marketing tools: 24/7 chatbots, sentiment analysis, fraud alerts and predictive insights.
AI text and image generation inside the Umbraco editor, plus reliable SCORM result tracking, for an e-learning platform on Umbraco Cloud.
AI hints that teachers approve before students see them, one sortable course list across databases and a sign-up flow that fits solo learners.
Quotes from verified Clutch reviews by Lokad and KursyPOL, whose TutorPRO MVP included an AI and OCR tool for creating exercises.
Given that the client (us) invest in giving them adequate feedback, they deliver high quality, maintainable code.
OSKI’s solutions are designed with clean code and attention to detail
The project team demonstrated a high level of commitment.
The best CRM for customer service depends on whether support needs its own help desk or a service module inside your sales CRM. We compare seven tools by ticketing, SLAs, omnichannel and AI pricing, then show when custom fits.
The best CRM for financial services depends on your core system, your recordkeeping duties and your team size. We compare six options for banks, credit unions, wealth firms and dealmakers by data model, integration and price, and show when to build a layer.
An AI chatbot costs from about $4,200 for a simple bot over a small set of documents to $136,000+ for a multichannel bot that acts in your systems. We break it down by scope, component and rate band, and compare running costs.
Anything missing? Send us your question and an engineer will answer, usually within one business day.
Contact usAssistants inside your platform, chatbots for customer requests, pipelines that transcribe or summarize audio, video and documents, text and image generation in a CMS, and machine learning for sentiment, fraud alerts and forecasts. Each of these is something we have shipped for a client, and the case studies on this page show where.
We have shipped projects on the OpenAI API, Azure OpenAI, Google Vertex AI with Gemini Pro, Flash and Flash-Lite, and Stability AI for images. For classic machine learning we have used TensorFlow, PyTorch, scikit-learn and AWS SageMaker. We choose the model per task, and one system often uses more than one.
We do not trust raw output. Depending on the task, generated code is run against the real system before anyone sees it, a person approves the output first, or the model fills a fixed schema that is checked afterwards. In the Lokad assistant, 82% of generated scripts ran on the first try and 94% by the second, because failures were caught and retried before reaching the user.
We can set up model access in your own cloud and provider accounts, under your API keys, so you control the terms and settings that apply to your data. Files, results and conversation history can stay in your infrastructure. For Lokad, conversation memory lives on Lokad’s own Azure event store.
By sending each job to the cheapest model that does it well enough, and by measuring cost per job from the first prototype. In the lecture transcription pipeline, choosing between Gemini Pro, Flash and Flash-Lite by task cut AI and API costs by up to 60%.
Yes, and we usually recommend it over a separate AI tool. We have connected AI services to existing CRM, ERP and marketing systems, added OpenAI and Stability AI generation to the Umbraco editor, and built an assistant directly into a client’s own platform.
Yes. Our engineers use AI tools in day-to-day delivery. By our own estimate this makes delivery about 3x faster. That is our figure, not an independent measurement, and every change still goes through code review and testing.
A spec, a Figma file or three paragraphs is enough to start.
Cookies on oski.solutions
We'd like to use analytics cookies to understand how the site is used, and marketing cookies to measure our ads. They are off unless you allow them. You can change your choice any time via “Cookie settings” in the footer. Cookie policy
