Hataf Atif

─ The planner was reverted before it was rebuilt

An LLM planner went in, came straight back out in favour of a deterministic placeholder, and was only then rebuilt properly as a standalone Chapter Planning Engine with explicit planning principles in the prompt. Shipping the intelligent version early would have meant debugging generation quality and scheduling logic at the same time.

─ Holidays and exams are inputs, not exceptions

Class paths are buffer-budgeted and re-plannable, so a two-week exam period changes the plan instead of invalidating it. Modelling lost time as a first-class input is what makes the plan survive a real school calendar.

─ Generation is per-slot and on demand

Lesson plans and formative assessments are generated for the slot a teacher is actually looking at, rather than batch-generating a term nobody reads. Cheaper, and it means a re-plan doesn't invalidate a pile of finished work.

─ Tenancy is an org API key, and that cut both ways

Per-org keys keep regional teams isolated, but a stale key surviving signup/login leaked across tenants until it was explicitly cleared on both paths.

─ One source of truth for training rules

Levels, attempt ceilings, pass marks and cooldowns resolve from a single place rather than being restated in the bot, the portal and the dashboard. Three copies of a pass mark is three chances to disagree with itself in front of a teacher.

─ A chat thread has to behave like a gradebook

Attempts, terminal states and cooldowns need to be strict even though the interface is a conversation. A level with no exam still needs a terminal state, or a teacher gets stuck in a lesson that never ends.

─ Certificates are minted on demand

PDFs are generated the moment they're earned and delivered in-thread, rather than pre-generated or emailed later. The reward has to land inside the same conversation that produced it.

─ Questions are pooled and varied per attempt

Fewer questions served per attempt, drawn varied from a pool — so a retake isn't a memory test of the first attempt.

─ Route point lookups instead of iterating

slack-search was rewritten so a direct lookup goes straight to the message rather than walking search results. About 20% fewer tool hops at equal correctness — the win came from removing a reason to guess, not from adding capability.

─ Per-team tokens over a central credential

Every team holds its own token. Slower to set up once, but it removes the shared secret and the person who owns it from the critical path.

─ Label placement is a solvable problem

The worldmap skill de-collides overlapping labels and picks label contrast per region automatically, so output is legible without a human nudging text boxes.

─ Extraction as a separate serverless function

POS and NER ran as their own endpoint rather than inside the pipeline, so language-model work could fail, retry and be redeployed without touching the ETL.

─ Websockets for the inbox, not polling

Support agents needed messages to arrive, not to be fetched. A serverless websocket endpoint kept that realtime without a server to babysit.

─ Provision the schema, don't document it

Setup creates the full Notion database schema itself. A README describing the columns you're supposed to add by hand is a README that goes stale in a week.