BackContext is live: one context layer for every agent you run.
Explore BackContext
Hytech Seeds takes orders as emails, PDFs and phone notes. They are full of variety names, lot codes and regional abbreviations that were never written down in one place. General-purpose models misread them, so the order desk retyped orders by hand, at about 6 minutes each.
Applied Crew connected BackContext to the order desk and its agent tools. It captured 206K order records and verified each one against what actually shipped, so the training data reflects real orders, not guesses.
We fine-tuned two open-weight models. hytech-orders-v2 extracts the order details from each email, PDF or phone note. hytech-catalog-v1 resolves the names and abbreviations customers use to the right variety and SKU.
On a held-out set of Hytech Seeds orders, tested side by side with Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra and GPT-6.1 Sol, hytech-orders-v2 reached 96.7% extraction accuracy. Claude Opus 5.5 reached 91.8% on the same orders.
The models went live behind a quality gate with frontier fallback. Order entry time dropped from 6 minutes to under 40 seconds, and the desk now reviews orders instead of retyping them. Every shipped order goes back through BackContext into the next version.
Hytech Seeds runs two models of its own in production. They read orders more accurately than Opus 5.5, they know the catalog the way the team does, and at Hytech’s volume the switch is projected to save $310K a year.
“Our order desk used to retype every email. Now the model reads them the way our team does, including the abbreviations nobody documented.” — Director of Operations · Agriculture
Hytech Seeds is a seed distributor whose order desk handles orders arriving by email, PDF and phone.
Seed distribution
hytech-orders-v2
+
hytech-catalog-v1
206K
Under 40 seconds
