Inspiration
All of our team members have participated in research at the university level, so we have first hand experience of the inner workings of a lab. Reflecting on our time working in labs, we came to an understanding that lab machinery and reagents contribute more to pollution than we realized. Identifying this correlation, we were inspired to develop a website that lab members could use to plan their week to limit energy usage and reagent waste, aiming for a sustainable solution.
What it does
Our website parses weekly calendars of each lab member with an uploaded pdf of the experimental protocol they have to carry out. These entries allow our workflow, composed of a protocol database and reagent hazards grouped by severity to identify overlapping times between labmates to create physical schedules for each person to follow. These schedules allow materials and energy to be saved on a scale impactful for sustainability.
How we built it
We match this input to a pre-existing curated protocol library using a Gemini API. We use the EPA datasets (part of the datasets given by DataHacks) for hazard classification and waste segregation of reagents, as well as a pre-existing catalog of lab equipment, to send a structured json to an overlap engine. This engine in turn uses custom algorithm heuristics to match the overlap reagents and equipment which are sorted by the EPA hazard priority. A Gemini API layer then explains this output to the user in natural language, where we can see an impact summary of the reductions made by this scheduling, a coordination list which tells us what experiments must be done together and how, and personalized calendars with the requisite procedures detailed for each individual lab member.
Challenges we ran into
Overlap sounds like one problem but it's three: reagents, equipment, and time windows each need their own rules, and a shared PCR only works if the thermal profiles match exactly. Cross-vendor naming was another challenge we encountered. Qiagen's Buffer AW2, Thermo's Wash Buffer 2, and NEB's 80% EtOH wash are the same reagent, but nothing in the protocol text admits it, so we read SDS sheets until the mapping stabilized. EPA data gave us trouble too since proprietary buffers aren't tracked individually, which meant classifying by component or by analogy and documenting every call. Carefully preparing and curating the libraries was an integral part of making our product effective for our users.
Accomplishments that we're proud of
Every hazard call on our site points to a real EPA entry. Paste any identifier into CompTox, TRI, or RCRA and it resolves. No LLM goes anywhere near a safety decision. We hand-curated the reagent map, waste compatibility matrix, and impact coefficients, and the schema is extensible, which is the point. The stack is split on purpose: Gemini handles the fuzzy job of turning PDFs into structured JSON, and a deterministic engine handles everything else, because LLMs cannot in their current state be reliably used to inform high-stakes chemical decisions.
What we learned
Greedy scheduling turned out to be a highly effective tool. It progressed through walking the week in priority order, merging the highest-impact overlap first, re-checking constraints, and moving to the next. While not entirely optimal, it is applicable for a realistic lab week, easy to reason about, and effective in achieving our desired outcome.
What's next for GreenBench
We hope to integrate more experimental protocols, curated faster now that Gemini can do the first pass on vendor PDFs and we validate against the schema we've already built. A newer update in the future for multiple labs with shared spaces, refrigerators and equipment. Dollar-cost savings alongside CO₂e, because reagents and electricity are expensive and labs move on money faster than they move on carbon. Per-step stability modeling so the engine understands an enzyme on the bench has a different shelf life than one on ice. And eventually cross-lab coordination and department-level benchmarking, because once people can compare their footprint to the lab next door, behavior actually starts to change.
Log in or sign up for Devpost to join the conversation.