Embedded Files
AI applications, agents & working prototypes
AI applications, agents & working prototypes
Selected projects connecting language models with business software, documents, images, video, and structured data. Each starts with a practical question and turns it into something that can be tested and improved.
61 completed Upwork jobs · 100% Job Success · Expert-Vetted · $400K+ earned
Upwork figures as of September 25, 2026. Completed jobs span consultations and other engagements.
Conversational sales & registration
Conversational sales & registration
Built agents that gather customer details and request a prefilled checkout link from an existing backend. A separate LLM combines that result with conversation history and business information to write the reply.
Contribution: agent and response workflows within a team product. Customers complete checkout themselves.
Customer messages → tool requests → checkout links and contextual replies.
Financial document processing
Financial document processing
Combined vision models, OCR, and LLMs to extract records from checks, money orders, and bank statements. Separate interpretation and formatting stages produce structured JSON.
Validation and reconciliation checks flag missing transactions, inconsistent amounts, and credit/debit errors for review.
Scans and document images → structured records and validation findings.
Procurement research agents
Procurement research agents
Created agents that search for equipment and suppliers using part numbers and descriptions. They collect prices, source URLs, and supplier contact details in one research workflow.
The result brings source-linked findings together for a purchasing decision.
Part numbers and descriptions → supplier research with prices and source links.
AI evaluation & product analytics
AI evaluation & product analytics
Built Python and SQL tools to review chatbot conversations and customer outcomes, plus LLM-assisted checks of FAQs and product catalogs.
Conversation-review tools and reports help account managers identify configuration problems and prioritize improvements.
Conversations and business records → quality findings and prioritized reports.
Flexible data normalization
Flexible data normalization
Developed AI-assisted workflows that map varied spreadsheet and transaction formats into consistent schemas.
The work handles differences in column names, table structures, and transaction formats so downstream software can use a wider range of inputs.
Inconsistent CSV and spreadsheet records → standardized application data.
Floor-plan understanding
Floor-plan understanding
Combined SAM 2 image segmentation with a vision LLM to identify rooms. Masks and bounding boxes locate regions; coordinates and grid markers give the model spatial context.
The model uses that context to connect room labels with specific locations on a drawing.
Floor plans, masks, and coordinates → room labels linked to locations.
Iterative image & sprite generation
Iterative image & sprite generation
Built a workflow linking image generation, visual critique, and editing. A critic LLM returns JSON with visual issues and editing instructions.
An editing agent uses those instructions and the previous image to create a revision, then submits it for another critique. Outputs include transparent PNG sprites, palettes, metadata, and asset manifests.
Visual brief → generated image → critique → revised assets.
Multimodal gameplay analysis
Multimodal gameplay analysis
Built a local prototype combining gameplay video, computer vision, OCR, and vision-language models.
It produces timestamped events, structured analysis, synchronized replays, and review dashboards. Supporting visual evidence remains available for inspection.
Gameplay recordings → evidence-linked events, analysis, and replays.
Page updated
Google Sites
Report abuse