preparing for jee over two years is fundamentally an engineering problem. you have 90+ advanced chapters across physics, chemistry, and math, thousands of module problems to solve, coaching lectures every afternoon, and test series every two weeks.
most students don’t fail because they lack intelligence. they fail because the system breaks down. they lose track of prerequisites, let backlogs pile up silently, or burn out trying to follow impossible schedules.
i tried using existing tools to manage the workload. every single one failed in practice.
the failure modes
first, i tried standard ai chatbots like chatgpt and claude. the problem is context amnesia. you can spend an hour setting up a study plan, and by the next morning the model has completely forgotten your past test scores, what module questions you solved, and where your actual weaknesses are. it invents a brand new, disconnected schedule from scratch every single day.
second, i tried database apps like notion and todoist. the issue here is logging friction. after an exhausting six-hour coaching day, the last thing anyone has the energy for is opening a database and filling out six properties across multiple rows. after two weeks of discipline, the logging fatigue wins and the system gets abandoned.
third, i tried study planner apps. almost all of them rely on fantasy scheduling. they budget time based on arbitrary question targets rather than empirical solving speeds. if you get sick or have a school exam and miss a single day, they dump all the missed work onto the next day, creating an exponential backlog avalanche that makes you want to quit.
i realized i needed a system that understood real preparation dynamics. so i built gale, named after gale boetticher from breaking bad for his obsessive precision and clean lab notes.
the core rule: the ai is not the memory
the foundational design principle behind gale is simple: the ai is not the memory; the ai operates the memory.
treating an llm as a persistent database is a mistake. factual academic data (chapter statuses, question completion tallies, accuracy rates, and error logs) should live in human-readable, structured files locally on your own machine.
the ai agent acts strictly as an operator. it reads the local state, calculates constraints, reasons through daily priorities, and updates files deterministically.
designing for zero friction
if a system requires high willpower to maintain, it will fail. i designed gale around a few practical rules:
1. conversational logging instead of manually updating tables, i just talk to the cli naturally:
“did 32 nlm questions from step 1, got 26 right, struggled with pulley constraints on inclined planes”
the agent parses the text, updates subject metrics, tracks accuracy percentages, and logs the concept trap in a single turn.
2. empirical capacity modeling gale schedules work based on physical reality. it uses actual available daily hours (differentiating school days, coaching days, and test weeks) multiplied by real solving speeds (like 10 to 14 mechanics problems an hour versus 18 to 24 chemistry problems an hour). no fantasy 14-hour daily targets.
3. anti-avalanche recovery when life disrupts the plan, gale absorbs the drift. if a day is missed, it adjusts priorities rather than doubling the next day’s workload.
what’s next
gale started as a personal tool to survive my own preparation. but the more i refined it, the more obvious it became that every serious student is dealing with the exact same broken workflows.
i am currently running a private beta and working on turning the core engine into a standalone product.
if you are building in the local-first or ai agent space, feel free to reach out at hello@h4rman.com.