Trade finance teams review letters of credit and guarantees that run past 100 pages. Manual review took a week per case, and text-based retrieval was undermined by OCR errors.

A production system that retrieves over document layout and visual structure rather than extracted text, extracts fields with a vision-language model, and runs agentic workflows over the results. It is served on GPU infrastructure and integrated with the bank's enterprise systems.

1. 01

   Ingest

   Letters of credit and guarantees, 100+ pages, as page images.

2. 02

   Embed

   Embeddings over layout and visual structure, not OCR text.

3. 03

   Retrieve

   Vector search and multimodal RAG over the document pages.

4. 04

   Extract

   Vision-language model field extraction.

5. 05

   Act

   Agentic workflows over the extracted results.

6. 06

   Integrate

   REST APIs and a .NET application layer into the bank's systems.

• Vision-language models
• LLMs
• Multimodal RAG
• Vector search
• Agentic workflows
• GPU inference
• .NET
• REST APIs

• 01Documents of 100+ pages per case
• 02OCR errors corrupting text-based retrieval
• 03Inference time and concurrency on shared GPU instances

• 01Document review cut from 1 week to 3 minutes
• 02Retrieval accuracy improved by 80%, OCR errors eliminated
• 03Inference time from 20 minutes to 40 seconds (30x) via GPU efficiency work and token caching
• 0420 concurrent users per instance with no measurable slowdown

Building automation workflows by hand means learning a builder and wiring every step yourself.

Users chat with an agent that builds their automation workflows on a canvas, and can drive the same agent over Telegram to do nearly the same thing.

1. 01

   Chat

   Conversational agent on the web or over Telegram.

2. 02

   Generate

   The agent builds the workflow through OpenRouter models.

3. 03

   Canvas

   Workflows rendered and edited on a React canvas.

4. 04

   API

   Hono and tRPC backend in TypeScript.

• TypeScript
• Hono
• tRPC
• React
• TanStack Router
• TanStack Query
• OpenRouter
• pnpm

• 01Live in production

Incoming documents arrive unlabelled and need classifying before any downstream processing.

A fine-tuned BERT classifier fed by OCR extraction with tailored pre-processing.

1. 01

   Extract

   OCR text extraction.

2. 02

   Process

   Tailored pre-processing of the extracted text.

3. 03

   Classify

   Fine-tuned BERT classifier (TensorFlow).

• Python
• TensorFlow
• BERT
• OCR

Detecting crops in field imagery as they change appearance through their growth stages.

A crop classification model used for feature extraction in object detection, with image segmentation across growth stages.

1. 01

   Classify

   Crop classification model.

2. 02

   Detect

   Classifier reused as the feature extractor for object detection.

3. 03

   Segment

   Image segmentation across growth stages.

• Python
• TensorFlow
• PyTorch
• Computer vision

• 01Led a team of 3 AI engineers, owning delivery, technical direction and architecture decisions across the team's projects.
• 02Owned the trade finance system after go-live: monitoring, incident resolution and iteration.

• 01Drove the trade finance solution through to production for a major bank in Kuwait, across letters of credit and guarantees in 100+ page documents.

• 01Cut document review from 1 week to 3 minutes.
• 02Reduced inference time from 20 minutes to 40 seconds (30x) through GPU efficiency work and token caching.
• 03Sustained 20 concurrent users per instance with no measurable slowdown.

• Open-source LLMs
• GPU inference
• Token caching
• RAG

• 01Built the embedding, vector-search and retrieval pipelines behind the trade finance product.
• 02Integrated with the bank's enterprise systems over REST APIs and a .NET application layer.

• 01Designed and built a production RAG system that retrieves over document layout and visual structure rather than extracted text.

• 01Improved retrieval accuracy by 80% and eliminated OCR errors entirely.

• RAG
• Embeddings
• Vector search
• REST APIs
• .NET

• 01Deployed hundreds of open-source models across multiple Linux GPU servers with varying hardware configurations, handling provisioning, serving and scaling.

• 01Built RAG applications with reasoning capabilities on top of the deployed open models, including the embedding and retrieval pipelines behind them.

• Open-source LLMs
• Linux
• GPU servers
• RAG
• Embeddings

• 01Developed text classification models for banking institutions.

• 01Built LangChain applications leveraging OpenAI GPT models.

• Text classification
• LangChain
• OpenAI GPT