Foodwit

Inspiration

The food and beverage industry spends $15B annually on compliance, $6.5B of which is due to recalled product. 4 of 5 root causes for recalled product is due to human error, a problem that can be mitigated with AI.

What it does

Ingests unstructured data (e.g., .pdfs, .doc, .xlsx) for raw material specifications (e.g., flour, sugar, butter) and recipes for finished food products (e.g., chocolate chip cookie). Passes information through a reasoning tool augmented with pertinent FDA information. Answers user generated prompts related to ingested data (e.g., what allergens are present in this product?)

How we built it

For the purposes of the hackathon, we used AutoGPT AI Agents to find and stich together Python libraries for document ingestion, parsing, and base algorithms. GPT3.5 was used as the base LLM, and AutoGPT AI Agents were used to develop and execute sub-tasks autonomously. Because GPT3.5 was used, agents were instructed to crawl the internet (up to date FDA regulations) to augment the model's base knowledge. Autonomous agents were used to field known use case prompts to simulate questions posed by a user. Specific compliance tasks for specific ingested data attributes were returned.

Challenges we ran into

Foundational building blocks are available, but token size limitations were consistently observed. A work around for this sprint was to use AI agents to develop rule-based algorithms to grab pre-processed data. The fundamental problems were formatting issues and token size. Eventually, we believe this can be resolved by using off the shelf OCR, parsing and human in the loop (HITL) quality scoring tools combined with a vector DB to orchestrate more efficient interactions between parsed and organized data and the augmented LLM.

Accomplishments that we're proud of

Our hackathon sprint suggests that data ingestion, OCR, parsing and quality scoring tools are largely available off the shelf, and together with human in the loop user oversight gives us a high degree of confidence. We also are proud of being a relatively non-technical team using AutoGPTs to rapidly develop a compliance tool. When combined with well established non-AI/ML tools, we feel confident in our ability to build and execute a suite of compliance tools.

What we learned

Token limitations are a hurdle everyone seems to be facing. Lighter versions of an LLM would be appropriate for a specific domain area of expertise. Tools for data ingestion, parsing, and quality automation for unstructured data are available, but interconnectivity between LLMs and peripheral data inputs require deeper orchestration and domain expertise.

What's next for Foodwit

Our Company is a regulatory and food safety services firm. We are a shortcut to ship a software product because we are domain experts, have a unique history of work product, established client relationships, and a healthy services business that can be used to continue to train and optimize a product. Most importantly, we have established trust with the food industry, which will be critical to adoption of this type of tool. Next steps are to fundraise, hire engineering talent, and change the way the industry works.

Built With

• autogpt
• gpt-3.5
• python