Work we're proudto put our name on.

What we built

Document AI ingestion: a vision model reads each page together with its text layer, Azure AI Document Intelligence reads the price tables, and the pipeline derives single products, their variants and their pictures from the raw files. Natural-language query understanding: a large model classifies the request, tags every part of it with a role from a closed set, and deterministic code resolves the tags into filters and ranking, so the model never adds what the user did not write. Hybrid retrieval: full-text and vector search merged by reciprocal rank fusion. Retrieval-augmented answers to questions, cited to the source page. A regression suite of real requests, run against the live model before any change ships. The whole on Azure serverless with a Python pipeline and API, an ASP.NET Core MVC site and an Angular administration application.

Highlights

LLM query interpretation with code-side resolution · Vision and document AI over raw PDFs · Hybrid FTS5 + vector retrieval · RAG with cited sources · Deterministic results from stored model readings · Model-in-the-loop regression testing · Python · .NET 10 MVC/Razor · Angular · Managed identities only

Industry

Furniture · Interior design · B2B sales

Technologies

What we built

Retrieval-augmented generation as a product: a document pipeline that turns PDF, Word, PowerPoint, Excel, CSV and text into Markdown, chunks it by section, enriches and embeds it, and stores versioned knowledge-base builds that are switched live atomically. Hybrid retrieval: vector similarity (normalized embeddings scored with SIMD dot products) fused with keyword search by reciprocal rank fusion, then re-ranked by a model. Conversation-aware answering: a small model rewrites each turn into a standalone question, a large model answers from the retrieved passages under a composed per-tenant system prompt, and a second check names any sentence the documents do not support before the answer is streamed. A knowledge-gap log that tells each customer what visitors asked and the documents did not cover. Around it, the full platform: tenant isolation with per-assistant storage, keys and origin rules; three embed surfaces (floating widget, in-page embed, hosted chat page) streaming over server-sent events; an admin application for accounts, plans, models and prompts; Stripe subscriptions with trials; and an SEO-first public site with a blog. An answer regression suite runs against the live API before every deployment.

Highlights

RAG grounded in the customer's documents · Hybrid vector + keyword retrieval with RRF and re-ranking · Standalone-question rewriting · Model-checked answers, no hallucinated sentences · Multi-format document ingestion · Multi-tenant SaaS with Stripe billing · Four applications: .NET MVC site, .NET Functions API, Angular portal, Angular admin · API Management gateway · Azure serverless

Industry

SaaS · Conversational AI · Small and mid-sized businesses

Technologies

What we built

A custom content management system in three applications. A server-rendered ASP.NET Core MVC website on .NET 10 whose pages are carved into several hundred editable sections, each read from Azure Table Storage with the written copy as its fallback and sanitized before it is shown. A custom blog built for search engines: posts and topic pages served as server-rendered HTML at readable addresses, a title and description on every page, a generated sitemap, drafts and publishing, and images converted to WebP on upload. A contact form that records every inquiry first and e-mails second, so that a mail outage never loses a message, and security headers on every response. An administration tool in Angular 22, signed in through Microsoft Entra ID with MSAL, its screens guarded by application roles, with a rich-text editor for pages and posts. Its API in C# on Azure Functions (.NET 10 isolated worker), which validates the user's token, checks the role each operation requires, and calls the website's own authenticated API, so that only one application ever touches storage. Settings are validated at startup, expected failures are results rather than exceptions, and the full detail of a fault goes to Application Insights while the visitor sees a plain message with a reference. The next stage, now in progress, is a custom AI-supported customer management system, including subscription management and document management.

Highlights

Client edits all content, no deployment, no developer · Custom content management system, data-driven pages · SEO-friendly: server-side HTML rendering, readable addresses, sitemap · Custom blog with topics, drafts and publishing · Three-tier chain: Angular admin, Functions API, MVC site · Entra ID sign-in with application roles end to end · Embedded AIChatPlacement assistant · In progress: AI-supported customer management system, subscriptions and documents · Four-language audience

Industry

Education and career services · Credential evaluation · Professional licensing

Technologies

What we built

A single-purpose landing page (the offer, the quote form, trust and product sections, a closing call to action) with a thank-you page, and a small API that validates each request and stores the lead; both deployed from GitHub.

Highlights

Quote form with event date, request type and an optional message · Leads stored in Azure Table Storage for the client's follow-up · Product cards worded for search engines, with the event types each rental serves · Mobile layout · In progress: starting prices, ad tracking (Facebook Pixel, Google Analytics) and a custom domain

Industry

Event rentals · Local services

Technologies