99%
Reduction in the effort of entering transcript data by hand
Built a transcript extraction service on Claude that reads any school's layout, returns structured per-course data, and takes manual data entry off teachers' desks for a student credit management platform.
Reduction in the effort of entering transcript data by hand
Extraction accuracy on formats never seen before
Time to turn a transcript PDF into verified, structured data
Industry: Education technology
Company size: Medium
Product: Claude Sonnet 4.6
Location: California, United States

Nothing on the platform works until a student's transcript is in it, and teachers were the ones entering it manually. As the client added districts, the backlog of transcripts waiting to be typed in grew with them.
For every student, a teacher opened the transcript PDF and hand-keyed each course row into the platform: the school, course, grade, term, credits attempted, credits earned. A single transcript spanned dozens of rows across pages and years, so teachers lost hours on work that added nothing to instruction.
Every district orders columns differently, notates grades differently, and encodes semesters differently. The old service was trained on four formats, so adding a fifth meant collecting samples, annotating them, and training a new model before a single transcript could be read. Growth kept hitting a wall.
Course placement, credit recovery, graduation eligibility checks, and state reporting all sat behind a task done by hand, one student at a time. A swapped credit value or a misread grade did not remain contained; it became an inaccurate picture of a student's progress toward graduation.
Simform built a transcript extraction system on Claude Sonnet 4.6 that reads any school’s transcript layout and returns structured per-course data through a single API call. Where a model trained on known formats can only handle those it has already seen, Claude reads each transcript fresh, so the service handles layouts it was never set up for.
Simform passes the raw transcript PDF to Claude Sonnet 4.6 as a base64 document block, so the model reads visual layout and text together in a single pass. Column alignment, multi-page structure, and layout cues stay available, with no separate OCR stage.
One system prompt carries all extraction rules, covering fields such as school, grade, and term; the semester-term distinction; the credit mapping; and an instruction to return null for any value not on the page. It runs at zero temperature, so a transcript yields a single answer.
Simform retired the custom-trained OCR model, the smaller refinement model, and the regex normalization rules that sat behind them. The service now returns a structured JSON array, one object per course, from a single API call.
Every extraction runs through automated test cases built from the same rules the prompt enforces, checking the JSON structure, the required fields, and the null handling for missing values. The structured output is verified against the defined schema before it reaches the platform.
This platform's backlog used to grow with every student enrolled and every district signed, since each one meant more transcripts waiting on manual entry. Now, new district transcripts are read by the same prompt on day one, so growth adds students to the platform without adding work to the queue.
Hand-keying every course row for every student, across dozens of rows and several school years, fell by 99%. A transcript that once had to be typed in one row at a time is now uploaded and read in a single pass, and teachers returned those hours to the classroom, where their time actually supports instruction.
The service used to read four transcript layouts, and reaching a fifth meant collecting samples, annotating them, and training a layout-specific model, a cycle that ran into days. A format nobody has seen is now handled by the same prompt at better than 95% accuracy, so the service reaches every new district on day one.
What Claude returns is what the transcript actually says. If a value is not on the page, the field returns null, so every grade and credit on the platform traces back to a real mark on a real transcript. A student's academic history reflects their record, not a model's best guess.
Hiren Dhaduk
Creating a tech product roadmap and building scalable apps for your organization.
We do not collect any information about users, except for the information contained in cookies. We store cookies on your device, including mobile device, as per your preferences set on our cookie consent manager. Cookies are used to make the website work as intended and to provide a more personalized web experience. By selecting ‘Required cookies only’, you are requesting Simform not to sell or share your personal information. However, you can choose to reject certain types of cookies, which may impact your experience of the website and the personalized experience we are able to offer. We use cookies to analyze the website traffic and differentiate between bots and real humans. We also disclose information about your use of our site with our social media, advertising and analytics partners. Additional details are available in our Privacy Policy.
These cookies are necessary for the website to function and cannot be turned off.
Under the California Consumer Privacy Act, you may choose to opt-out of the optional cookies. These optional cookies include analytics cookies, performance and functionality cookies, and targeting cookies.
Analytics cookies help us understand the traffic source and user behavior, for example the pages they visit, how long they stay on a specific page, etc.
Performance cookies collect information about how our website performs, for example,page responsiveness, loading times, and any technical issues encountered so that we can optimize the speed and performance of our website.
Targeting cookies enable us to build a profile of your interests and show you personalized ads. If you opt out, we will share your personal information to any third parties.