The problem nobody sees on the shelf Every packet of chips, tray of tomatoes or bag of spices on a shelf is the result of a packaging decision. That decision quietly determines how long the food stays fresh, how much it costs to make and ship, and whether the packaging ends up recycled or in a landfill. Getting it wrong has real consequences: spoiled produce, stale snacks, wasted money and avoidable plastic waste. Yet choosing the right material is not simple. A fresh fruit is alive and keeps respiring, so it needs a film that breathes. A fatty snack needs a very strong oxygen barrier to stop it going rancid. Frozen meat needs puncture resistance. Each choice involves trade-offs between barrier performance, sealability, strength, cost and sustainability, and small food businesses often lack a packaging engineer to guide them. PackAI was built for the to close that gap.
Figure. The PackAI sign-in screen.
What PackAI does
PackAIis a webapplication that turns a handful of food and storage parameters into a complete packaging recommendation. The user signs in, picks a commodity category (fresh fruits, vegetables, snacks, grains, bakery, meat, dairy or spices) and a specific food item, then enters the properties that matter: •Moisture content and oil or fat content, as percentages • pH value and respiration rate •Desired shelf life, storage temperature and relative humidity •Transportation condition and storage type (ambient, chilled or frozen) When exact values are unknown, a built-in reference chart lists typical ranges for 17 common foods, and a Use button auto-fills the form.
****Figure.The reference chart of typical food and storage parameters
With one click on “Analyze & Recommend”, PackAI returns a structured
result. It names the recommended material, for example micro-perforated film for fruit, metallized PET/PE for snacks, PA/PE vacuum packaging for meat or an aluminium foil laminate for spices, together with a match score. It also gives the technical profile: oxygen and water vapour transmission levels, film thickness, sealability, strength and suitability for modified atmosphere packaging (MAP). For MAP it proposes a gas mix of oxygen, carbon dioxide and nitrogen.
Figure 5. The recommendation for potato chips: a 97% match with MAP gas mix, sustainability, cost, shelf life and recyclability.
The sustainability lens
What sets the results apart is that they do not stop at performance. Each recommendation carries a sustainability score, a cost level, an estimated shelf life and an honest recyclability verdict. Most high-barrier laminates are only “conditionally recyclable”, and PackAI says so, with a note that recyclability depends on local collection and sorting. It then lists greener alternatives such as mono-material PE or PP films and barrier-coated paper, plus example manufacturers such as Amcor, Unflex, Mondi, Huhtamaki and Berry Global.
How the recommendation is built
The engine begin swith abase profile for each commodity category and then adjusts it to the user’s inputs. High moisture (above 60%) or high humidity (above 75%) tightens the water vapour barrier requirement. High fat content raises the oxygen-barrier requirement. Rough transport increases strength and thickness, frozen storage calls for a thicker structure, and a high respiration rate raises MAP suitability. Estimated shelf life also shifts with the desired shelf life and storage type. The logic is transparent, rule-based and predictable, which suits a domain where explainability matters more than a black-box score. Around this engine sits a small but complete full-stack application: a vanilla HTML, CSS and JavaScript front end, an Express 5 server on Node.js, and SQLite for user accounts and every saved recommendation, so users keep a history of their analyses.
Adding memory with Hindsight
The twist was persistent AI memory. Most tools treat each session as a blank slate: a user who cares about recyclability has to say so every time. PackAI integrates Hindsight, a memory service, through three operations that map neatly onto the workflow. •Retain: each recommendation, with its inputs and outcome, is stored as a memory. Users can also save explicit preferences such as “Prefer Recyclable”, lower cost or longer shelf life. •Recall: before a new analysis, PackAI retrieves relevant past memories and displays them in a “What PackAI Recalls” panel. •Reflect: the server also exposes a reflection endpoint for higher-level summaries of what has been remembered. A deliberate design rule keeps this safe: memory personalizes the presentation, never the engineering. If recalled memories mention recyclability, the greener alternatives move to the top of the list and the cost note is framed around that preference. The technical recommendation itself is not overridden. A memory status indicator in the interface shows whether Hindsight is connected, which makes the feature easy to demonstrate live.
Explaining it with Groq
These Cond layer is a live AI advisor. After each recommendation, PackAI sends the inputs, technical result and any recalled memories to a Groq-hosted language model, and streams the reply into the page word by word using Server-Sent Events. The default model is openai/gpt-oss-20b, with an automatic fallback to llama-3.3-70b-versatile if the first is rate limited, unavailable or returns an empty answer. A plain JSON endpoint serves as a backup if streaming fails. The prompt is tightly constrained. The model must treat the technical fields as the primary engineering limits, use memories only as preferences, never invent technical facts, and answer in four fixed sections under 180 words: AI Advice, Why This Fits, Memory Used and Next Step. This keeps the output short, consistent and clear about what is a fact and what is a remembered preference
Figure 6. Memory status, recalled memories and the streaming Groq advisor (emails blurred).
Engineering decisions worth noting
Result saved to SQLite and Hindsight Several choices make the project sturdier than a typical demo. API keys for Groq and Hindsight live only on the server and never appear in front-end code. Hindsight calls have an eight-second timeout, so a slow memory service cannot freeze the interface, and the memory layer is optional: without a key, the app and the Groq advisor still work. The Groq client handles reasoning-model quirks by keeping reasoning effort low and leaving token headroom for the visible answer. One-click launchers (START.bat and START.sh) create the environment file and install dependencies, so a judge can run the project in minutes.
Figure 7. Hindsight code snippet in VS Code.
A suggested demo story
The project is easiest to appreciate as a two-step story. First, the user logs in, clicks “Prefer Recyclable” and analyzes a food item. PackAI recommends a material and stores the memory. Second, the user analyzes a different product. This time PackAI recalls the earlier preference, surfaces recyclable alternatives first, and Groq explains the result while referencing that memory. The audience can see the assistant learning from a past interaction, which is the point of the memory layer.
Limitations and next steps
PackAIis a prototype, and it is worth being clear about what that means. The recommendation engine is rule-based with hand-set profiles, not a trained machine-learning model, so its outputs are guidance rather than validated engineering specifications. The sustainability scores and shelf-life estimates are indicative. Before real deployment, the following would be priorities: Hash and salt passwords, and add proper session tokens; the current login is a prototype-level implementation. Validate the material profiles and shelf-life ranges with packaging scientists and real test data. Replace the fixed rules with data-driven models, keeping the explanations transparent. Add regional recyclability data so the verdict reflects what a user’s local facilities actually accept. Let users rate recommendations, so the memory layer learns from outcomes as well as preferences.
Conclusion
PackAI showshow three ideas can work together in one small product: deterministic engineering logic for correctness, persistent memory for personalization, and a streaming language model for clear explanation. Each layer has a defined job, and none is allowed to overstep it. For food producers, especially smaller ones, that combination points toward a future where good packaging advice is quick to get, tailored to the user, and honest about its environmental cost.
PackAI uses Hindsight, an agent memory system, through three operations. The Hindsight docs cover the API, and Vectorize has a good explainer on what agent memory is.