AI-Assisted Development: Domain Expertise Trumps Raw Code Generation

A materials science graduate leveraged Anthropic's Claude to build a production FTIR analysis platform, claiming AI wrote 70% of the code. This case study highlights the critical role of domain expertise in AI-assisted development.

A materials science graduate, identified as Bob Lee, reports building a production FTIR analysis platform where Claude wrote 70% of the code. This platform, designed to search 135,000 FTIR spectra, addresses a gap left by expensive enterprise solutions and outdated Excel macros.

Lee, who describes himself as a domain expert with basic Python skills, details a multi-step process for developing the system. His experience suggests that while AI can accelerate code generation, human oversight and specialized knowledge remain indispensable for validating output and navigating real-world complexities.

Core Algorithm Development

The initial phase focused on the core FTIR spectrum matching algorithm. The founder claims Claude assisted with writing the peak-matching loop, setting up the Django project structure, and designing the database schema for the spectral library. This foundational work established the technical framework for the application.

However, Lee states he was responsible for critical domain-specific tasks. These included determining appropriate tolerance values for different wavenumber regions, validating match results against known materials, and rejecting initial algorithm designs that appeared correct on paper but failed with real data. The founder's assessment is that AI accelerates coding, but domain expertise remains the bottleneck for chemical accuracy.

Parsing Diverse Instrument Files

The most significant technical challenge involved parsing data from six distinct FTIR instrument file formats: SPA (Thermo Nicolet), SPC (GRAMS), OPUS (Bruker), CSV, JDX (JCAMP-DX), and XLSX. These formats range in difficulty from simple CSV to highly proprietary binary structures like OPUS.

Lee reports that Claude was instrumental in generating binary file parsers directly from format documentation, extracting peak tables, and handling edge cases such as missing metadata or non-standard headers. Despite Claude's proficiency, the founder identified and corrected "three subtle byte-offset errors" that would have silently corrupted data. This highlights the necessity of human review, even for AI-generated code that appears functional.

Integrating with AI Agents via MCP

Beyond traditional user interfaces, the platform incorporates an MCP (Model Context Protocol) server. This component, located at fastapi_server/mcp_server.py, allows other AI agents to directly call the FTIR analysis tool. Instead of human input through a web form, AI agents can send structured requests and receive structured results.

This integration points to a future where specialized tools are not just for human users but also for inter-AI communication, enabling more complex automated workflows. The founder's decision to implement an MCP server indicates an understanding of emerging AI interaction paradigms, moving beyond simple API exposure.

What We'd Change

The founder's claim that AI wrote 70% of the code, while compelling, lacks specific metrics for verification. This percentage likely refers to lines of code or initial scaffolding rather than the most complex or critical logic. Future iterations of this playbook would benefit from more granular data on AI contribution, perhaps tracking code blocks or function implementations.

Furthermore, the success of this project is heavily reliant on the founder's deep materials science expertise. A developer without this domain knowledge would likely spend significantly more time validating AI outputs, even with advanced AI assistants. The playbook would need to explicitly address how non-domain experts could acquire or simulate this critical validation layer, or acknowledge that this specific approach is best suited for expert-builders.

Finally, the reliance on a single AI model (Claude) introduces potential vendor lock-in. While effective for this project, a more robust architecture might explore model agnosticism or integrate multiple AI providers to mitigate risks associated with API changes, pricing fluctuations, or model deprecation. Diversifying AI tooling could enhance long-term maintainability and flexibility.

This project demonstrates a viable path for domain experts to build specialized tools, even with limited coding experience, by leveraging AI for code generation. The critical insight is that AI accelerates the how of coding, but human expertise remains paramount for defining the what and validating the correctness.

The investor read

This signal underscores the increasing efficiency of niche SaaS development, particularly when domain experts leverage AI. While the specific FTIR analysis market is small, the trend of AI-assisted development enabling non-technical founders to build production-grade tools is significant. Capital attention is shifting towards platforms that abstract away coding complexity, allowing experts to focus on problem-solving. For investors, this highlights opportunities in vertical SaaS where deep domain expertise is a competitive moat, even if the product itself is bootstrapped. The MCP server integration also points to a nascent but growing market for AI-native tools designed for inter-agent communication, potentially opening new infrastructure plays.