Built for an organization that needed to turn messy scanned reports into searchable, structured knowledge. The solution automated document processing and made large volumes of previously hard-to-use information easier to search, analyze, and reuse.

Self-hosted OCR + Vision-Language Model pipeline running on GPU with CUDA optimization. Handles scanned documents, complex layouts, tables, and embedded images. Benchmarked multiple OCR engines (PaddleOCR, Nanonets, MistralOCR, Deepseek OCR, OlmOCR...) against a representative corpus to select the optimal engine per document type. Includes intelligent chunking with semantic embedding for downstream RAG.

Built for a legal-services organization dealing with dense, non-standardized regulations. The system retrieves and explains relevant legal passages, helping users understand the right sections faster instead of manually searching through long regulatory documents.

Multi-agent RAG pipeline for complex legal document corpora. Custom AI chunker segments non-standardized legal text while preserving cross-references. Hybrid retrieval combines BM25 keyword search with vector similarity search, fused via Reciprocal Rank Fusion (RRF) for optimal ranking. LLM-as-Judge validation loop scores retrieved passages for relevance and completeness before answer generation. Metadata enrichment adds jurisdiction, document type, and date filters.

Built for a healthcare organization that needed to answer complex questions about coverage, procedures, and documentation. The solution helped users get accurate, source-backed answers without depending on support staff for every request.

Retrieval-based assistant for Health AI. Answers questions about health coverage, procedures, and documentation requirements by retrieving from official health documents - PPTX presentations, SQL databases, and regulatory text. Self-hosted infrastructure: Supabase for data and authentication, Qdrant for vector storage, OpenRouter for multi-model LLM access. Hugging Face embeddings with reranking pipeline. Arize Phoenix observability for tracing retrieval quality and latency. Chunked document retrieval with relevance scoring, source attribution on every answer, and memory for multi-turn conversations about complex healthcare procedures.

Built for an educational institution that needed to provide students and staff with reliable academic information. The assistant made schedules, exam details, and institutional content easier to access through a simple chat interface.

Multilingual retrieval-based assistant for a Serbian higher-education institution, fielding student and staff questions about exam schedules, deadlines, and general academic information. Hybrid retrieval splits factual exam queries to a live MySQL table so answers always reflect the current schedule, and routes everything else to semantic search over scraped institutional content (Pinecone, paraphrase-multilingual-MiniLM-L12-v2 embeddings, Gemini synthesis). Custom ingestion pipeline transliterates Cyrillic to Latin, strips HTML boilerplate, and chunks with RecursiveCharacterTextSplitter before embedding. Conversation history and authentication persisted in MySQL; Streamlit frontend for student-facing access.

Built for teams working with large and inconsistent document collections. The pipeline improved AI retrieval quality by turning messy documents into cleaner, more meaningful sections before they were used by downstream systems.

AI agent-driven document chunking pipeline that analyzes document structure and contextualizes chunks with logical boundaries. The agent determines optimal split points based on semantic meaning, section hierarchy, and content type. Falls back to LangChain splitters (RecursiveCharacterTextSplitter, MarkdownHeaderTextSplitter) when agent segmentation is not cost-effective. Cleans headers and footers, normalizes metadata for vector storage, detects tables, and generates image descriptions for embedded visuals. Built for heterogeneous, large-scale document collections where standard splitters break: mixed formats, messy structure, tables, visuals, and schema normalization across sources.

Built for a team developing AI question-answering systems. The framework helped compare models, prompts, and retrieval strategies so the team could improve quality with measurable feedback instead of guesswork.

Automated evaluation and testing pipeline for LLM question-answering, retrieval quality, and answer reliability. Built custom metrics alongside DeepEval framework and NLP algorithm metrics for comprehensive benchmarking of retrieval and generation quality. Golden datasets enable rapid comparison of prompting strategies, chunking approaches, and model selections across different providers. Continuous improvement pipeline that feeds evaluation results back into system tuning.

Built for an established German company selling technical products with detailed specifications, use cases, and installation requirements. The assistant helped customers find the right product, understand installation steps, and compare options more easily, reducing repetitive support questions and improving self-service.

Website assistant that answers questions about product catalog, compares multiple products side-by-side, and delivers detailed installation instructions - all in German. Collects contact details through guided flows with suggested questions to reduce friction. Analytics dashboard tracks conversations, user feedback, and conversion paths. FAQ system with confidence scoring to speed up self-service resolution.

Built for builders and architects working with complex construction regulations. The assistant retrieves and explains the right regulatory passages, helping users understand requirements faster and reducing the risk of relying on the wrong interpretation.

Specialized RAG system for building code and construction regulation documents. Parses legal text with structural awareness - chapters, sections, subsections, and amendments - to maintain hierarchical context. Retrieves exact passages with citation metadata (document, chapter, section, paragraph) so every answer points to the specific regulatory source. Source verification step cross-checks retrieved citations against the original document to prevent hallucinated references.

Built for an analytics-focused organization that needed easier access to both database records and written knowledge. The system allowed users to ask questions in natural language instead of writing SQL or manually searching through documents, while automatically generating charts and visualizations to make insights easier to understand.

Natural-language analytics interface combining structured SQL querying with semantic vector search. Routes queries between OpenAI and Google Gemini based on complexity. Supabase integration for data management, authentication, and real-time sync. Automated retrieval with explanation generation and analytics tracking dashboard.

Built for a platform that had no existing documentation. The system explored the application, extracted product knowledge from its screens and workflows, and generated structured user guides automatically. As the product changed, the knowledge base could be regenerated and kept up to date.

Autonomous crawling agent that structures platform documentation into a queryable knowledge base. Crawls unstructured sources, extracts meaningful content, and organizes it for team and customer access. Generates step-by-step instructions on how to use the web app based on documented features and workflows. Self-updating architecture that detects documentation changes and regenerates affected sections.

Built for internal teams that needed faster access to company knowledge. The query interface reduced time spent searching through documentation by returning specific answers with source attribution.

Query interface over generated knowledge bases. Delivers instant, specific answers from organizational data - not web search. Retrieves from collected company knowledge with source attribution. Designed for teams that need fast answers without hunting through documentation.

Built for environments where important information was spread across both text and visual materials. The system allowed users to ask questions and retrieve relevant answers from diagrams, manuals, and image-heavy documents.

Multimodal extension to the text-based assistant. Users ask questions in natural language and the system retrieves answers directly from a vector store of images - product manuals, installation diagrams, technical sheets - without OCR or caption generation. ColPali and ColQwen encode each image into late-interaction embeddings that capture fine-grained visual and textual features. At query time, the user's question is embedded and matched against this image vector store via similarity search, returning the most visually and semantically relevant pages. Answers are grounded in the retrieved image content combined with text documentation.

Built for a manufacturer with complex machine specifications and maintenance records. The system allowed workers to ask questions in plain language instead of manually searching through nested folders and technical files.

Natural language interface for deeply nested tree/folder structured data. Parses hierarchy of machine specifications, maintenance schedules, and operational status stored in folder structures. Plain-language queries return precise answers without manual navigation through nested directories.

Built for a service business that was missing calls and losing booking opportunities. The AI receptionist answered calls, guided customers through reservations, and reduced the need for staff to handle repetitive phone conversations.

AI that answers your phone calls, takes bookings, and handles the full reservation flow through natural conversation. Built on SIP with real-time speech processing, it integrates directly into your dispatch system - no hold times, no extra operators, no missed calls.

Built for a company that needed clients and employees to quickly understand its products, services, and internal policies. The assistant improved self-service and revealed common knowledge gaps through conversation analytics.

Retrieval-based assistant over company documentation - product specs, service descriptions, policies, and FAQs. Two response modes: concise answers for quick lookup, and expanded mode that retrieves additional context from related documents for deep dives. Auto-generated suggested questions based on retrieved context guide users to relevant follow-ups. Analytics track which questions are asked most and where the knowledge base has gaps.

Built for a fintech startup creating a more personal way for users to understand their money. The platform turned income, spending, goals, and financial questionnaires into conversational insights and weekly summaries.

Full-stack AI finance platform with WhatsApp as the primary user channel. Built and integrated the WhatsApp bot end-to-end: onboarding flow guides users through account setup, income categorization, and goal definition via natural conversation, and weekly AI-generated summaries are delivered directly to WhatsApp with spending breakdowns, trend analysis, and goal progress. AI layer analyzes financial profiles to generate personalized advice and runs calculations on income, expenses, and goals. AI-generated Pro Plans built by asking users detailed questions about their financial situation, risk tolerance, and objectives - producing structured, actionable financial roadmaps. Dynamic UI updates reflect personalized insights per user.

Built for a market-intelligence company that needed structured data from thousands of websites. The pipeline extracted pricing, features, and company metadata from more than 2,000 websites and turned messy public data into usable market intelligence.

Filtered a large candidate URL set down to usable sites before extraction. Crawl4AI as the primary scraping engine for high-performance, JavaScript-aware crawling at scale. Proxy rotation and browser automation handle difficult, JavaScript-heavy, and protected sites. Vision AI fallback (Gemini) processes sites that block even headless browsers. Pydantic schema enforcement ensures structured, consistent output regardless of source heterogeneity. Extracted pricing tiers, feature lists, and company metadata with precision.

Built for an event-planning company that needed to help users find suitable venues faster. The assistant guided users from requirements gathering to venue discovery, ranking, and personalized outreach.

Conversational AI using Google Places API for venue discovery. Follow-up questioning narrows requirements until search criteria are clear. Web scraping and reranking filter results. Generates personalized outreach emails for selected venues. End-to-end flow from discovery to contact in a single conversation.

Built for a mobile-app startup that wanted users to interact with AI through voice. The solution delivered a production-ready mobile experience with real-time speech interaction, authentication, and subscription management.

React Native mobile application with AI voice conversations. Vapi.ai handles real-time speech processing. Supabase provides authentication and data persistence. RevenueCat manages In-App Purchase subscriptions and monetization. Production-ready iOS build for App Store deployment with full subscription lifecycle handling.

Built for a biotech research team working on drug-discovery workflows. The pipeline helped automate complex molecular research tasks and made research papers and clinical data easier to retrieve inside one workflow.

End-to-end molecular simulation pipeline for precision medicine research. Automated protein-ligand docking with AutoDock Vina and GNINA. RAG system built on research papers and clinical datasets, integrated into CellType CLI for interactive research analysis. Proposed AxialBridge patient app for EULAR - ML/CV-based Axial Spondylitis detection from MRI and ultrasound imaging.

Built for teams that needed better answers to complex, multi-step questions. The system combined relationship-based reasoning with semantic search to retrieve more useful context than either method alone.

Neo4j knowledge graph combined with vector embeddings for hybrid retrieval. Outperforms vector-only approaches on multi-hop queries by following relationship paths. Outperforms graph-only approaches on semantic similarity matching. Best of both architectures for complex reasoning tasks.

Built for an experimental AI platform that needed more engaging voice conversations. The system routed users between different AI personalities and created a more dynamic spoken interaction experience.

LangGraph state machine with conditional edges for contextual routing between conversation personas. Real-time speech classification determines which agent handles each user turn. ElevenLabs TTS for natural voice output.

Built for teams customizing language models. The automation reduced failed training runs by validating files, estimating costs, and monitoring fine-tuning jobs before mistakes became expensive.

Automated SFT pipeline via API with pre-flight validation. Training file format checking prevents costly failed jobs. Cost estimation before submission. Real-time monitoring of training progress and evaluation metrics. Reusable system that protects against common fine-tuning mistakes.

Built for developers who wanted to experiment with model training locally instead of relying on cloud infrastructure. The pipeline made local fine-tuning, embedding generation, reranking, and benchmarking easier to run on Apple Silicon.

MLX-based pipeline for Apple Silicon: LoRA fine-tuning, embeddings generation, reranking, and contrastive learning. Core tasks benchmarked with MTEB for quality validation. High-throughput local training on Apple Silicon with BF16 precision. Self-hosted training without cloud dependency.