From Feature Shipper to AI Innovator
For five years, the author shipped features weekly at Tittri, a company serving a UK-based fintech client handling business loans. The feedback was clear: reliable, but reactive. The author was building exactly what was asked, end-to-end, focusing on business-critical web applications where the frontend was the operational engine. While effective, this wasn't enough. A shift was needed, from executing tasks to driving innovation.
In January 2026, the author decided to proactively pitch an AI integration. The target was the core product: a platform for business loan applications. This wasn't a theoretical exercise; it involved real users, real data, and real financial stakes. The initial presentation to the team and the client sparked immediate interest, raising expectations. The challenge was significant: no prior AI integration experience, no readily available tutorials, and the imperative to deliver tangible benefits to brokers. The journey from idea to production was about to begin, with no pre-written map.
Navigating the Uncharted: AI Integration Challenges
The author's approach to this project was pragmatic and iterative, driven by the lack of established patterns. The first hurdle was defining the AI's objective. Instead of aiming for a broad, complex system, the focus narrowed to a specific pain point: automating the initial assessment of loan applications. This involved processing unstructured data from broker submissions – a common bottleneck in traditional lending processes.
The core technical challenge lay in building a robust pipeline that could ingest, process, and analyze diverse data types. This meant moving beyond simple API calls or off-the-shelf solutions. The pipeline needed to handle document parsing (PDFs, scanned images), natural language understanding (extracting key financial figures and applicant details from free-text fields), and a scoring mechanism to flag applications requiring immediate human review versus those that could proceed with minimal intervention.
A significant hurdle was data preparation. Real-world fintech data is messy, inconsistent, and often contains sensitive personal and financial information. Cleaning, anonymizing, and labeling this data for training AI models required substantial effort. Unlike academic research, where datasets are often curated and clean, this project demanded a 'dirty data' approach, building robustness into the pipeline itself to handle imperfections.
The author adopted a phased development strategy. Phase one focused on building a minimum viable product (MVP) that could demonstrate core functionality: ingesting a loan application document, extracting key fields, and providing a basic risk assessment score. This MVP was crucial for gaining client buy-in and refining the project's direction based on early feedback. It also served as a testbed for the chosen technologies.
Technology Stack and Implementation Details
The technology choices were dictated by the need for flexibility, scalability, and integration with the existing Tittri infrastructure. Python emerged as the primary language due to its rich ecosystem of AI and data science libraries.
For document processing, libraries like PyPDF2 and OpenCV were employed for handling PDFs and scanned images, respectively. Optical Character Recognition (OCR) was a critical component, with the author likely leveraging tools such as Tesseract or cloud-based OCR services to convert image-based text into machine-readable format. This was particularly important for older or scanned application forms.
Natural Language Processing (NLP) was central to extracting meaningful information. Libraries like spaCy or NLTK were used for tokenization, part-of-speech tagging, and named entity recognition (NER) to identify financial figures, company names, loan amounts, and repayment terms. Custom regular expressions and rule-based systems were developed to handle the specific jargon and data formats common in business loan applications.
Model selection was guided by the problem's nature. For risk scoring, simpler, interpretable models like Logistic Regression or Random Forests were initially favored. These models offer transparency, which is crucial in a regulated industry like fintech. The author likely experimented with these before potentially exploring more complex models if performance warranted. The key was to build a model that not only predicted risk but could also provide a rationale for its assessment.
Deployment was another critical consideration. The pipeline was containerized using Docker to ensure consistency across development, testing, and production environments. Integration with the client's existing loan management system was achieved via a RESTful API built using Flask or FastAPI. This allowed the AI's insights to be seamlessly incorporated into the brokers' workflow without requiring a complete system overhaul.
Continuous integration and continuous deployment (CI/CD) pipelines were established to automate testing and deployment, enabling rapid iteration and bug fixing. Monitoring tools were integrated to track pipeline performance, model drift, and potential errors in real-time.
From MVP to Production: Iteration and Client Impact
The transition from an MVP to a production-ready system involved rigorous testing and refinement. User Acceptance Testing (UAT) with a select group of brokers provided invaluable feedback. They highlighted areas where the AI's interpretation was inaccurate or where the output format was not intuitive for their daily tasks.
One key area of iteration was the confidence scoring and alert system. Initially, the AI might have simply flagged applications as 'high risk' or 'low risk.' Through user feedback, this evolved into providing a more granular risk assessment with explanations. For example, instead of just flagging a loan amount as potentially too high, the AI might specify that it exceeds the typical loan-to-value ratio for the applicant's industry, citing specific data points from the application. This level of detail empowered brokers to make faster, more informed decisions.
The impact on the fintech client was measurable. By automating the initial screening and data extraction, the time required to process a loan application was significantly reduced. This led to faster turnaround times for brokers, improving their efficiency and client satisfaction. Furthermore, the AI helped identify potentially fraudulent applications or those with missing critical information earlier in the process, reducing operational risk and potential financial losses.
The author emphasizes that this was not a 'set it and forget it' project. Ongoing monitoring and retraining of the AI models are essential. As market conditions change, new loan products are introduced, or the nature of applications evolves, the AI needs to adapt. This requires a commitment to continuous improvement and a feedback loop between the AI system, the users, and the development team.
Lessons Learned and Future Directions
The experience at Tittri underscores several critical lessons for building production AI systems, especially in regulated industries:
- Start with a specific problem: Don't try to boil the ocean. Focus on a well-defined pain point where AI can provide clear, measurable value.
- Embrace 'dirty data': Real-world data is rarely perfect. Build pipelines that are resilient to inconsistencies and errors.
- Prioritize interpretability: In finance, understanding *why* an AI makes a decision is as important as the decision itself. Favor explainable models where possible.
- Iterate with users: Continuous feedback from end-users is crucial for refining the AI's performance and ensuring it meets practical needs.
- Plan for maintenance: AI systems are not static. Budget and plan for ongoing monitoring, retraining, and updates.
The author's journey from a feature-focused engineer to an AI pipeline builder demonstrates a successful pivot driven by initiative and a willingness to tackle complex, uncharted territory. This project not only delivered value to the fintech client but also significantly expanded the author's skillset and Tittri's capabilities in AI-driven solutions.
What nobody has addressed yet is how this kind of in-house AI development capability, born from individual initiative, can be systematically scaled and replicated across other clients or projects within Tittri without relying solely on individual engineers spotting opportunities.