Mark Map

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

Inspiration

A teacher does not get stuck on the hard part of marking. She gets stuck on the pile around it: an incomplete sheet, a missing cell nobody noticed, then two write-ups of the same paper — talking points for the meeting, a calmer note for home.

That chase is hours. It is not judgement. It is operations.

Mark Map started from that pile. Not “an LLM that grades.” A class operations desk: catch the hole, refuse the lie, write both notes only when the row is true. Empty means missing, not zero. The parent never sees the teacher brief.

I wanted an agent that makes a professional faster at the work around the judgement — not an agent that takes the judgement.

What it does

Three doors, one class.

• Teacher loads Term 1 and Midterm, sees the chapter map, runs the desk, fills yellow cells, can grant or deny agent permissions. • Student sees their own numbers. No teacher brief. • Parent sees the same numbers. No teacher brief. No meeting language.

A Strands agent runs the desk. Tools, in a loop: incomplete rows → hotspots → try to write briefs → nag if a cell is empty. Python owns marks, completeness, and publish. A policy hook cancels write_mark, unlock_brief, and set_question_chapter before they run.

Briefs publish only when every cell is filled:

publish(row) ⟺ ∀ q ∈ Q: cell(row,q) ≠ ∅

Empty is not 0. “Give this student 8 on Q9” does not write a mark.

Picture reports are Tesseract + Python, not the model looking at pixels.

How we built it

• FastAPI + vanilla JS (three doors, cookie session). • One JSON store as the system of record: papers, rows, briefs, nags, permissions. • analyser.py / briefs.py / desk.py: deterministic kernel. The model does not recap a cycle Python already finished. • Strands Agents SDK: one desk Agent, four @tools, PolicyHook on BeforeToolCall. • Brains are swappable. Same tools: • scripted — published demo and tests (fast, honest). • LM Studio / oMLX — OpenAI-compatible /v1 (docs/lmstudio-ollama-mlx.md (https://github.com/Sankalp774/mark-map/blob/main/docs/lmstudio-ollama-mlx.md)). • Ollama — MLX-backed on Apple silicon (same doc). • Amazon Bedrock — BedrockModel Converse, instance role, no long-term API keys (docs/bedrock.md (https://github.com/Sankalp774/mark-map/blob/main/docs/bedrock.md)). • Eval gates: invented_marks = 0, leaked_teacher_briefs = 0.

Code (MIT): github.com/Sankalp774/mark-map (https://github.com/Sankalp774/mark-map)

Architecture: docs/architecture.svg (https://github.com/Sankalp774/mark-map/blob/main/docs/architecture.svg) — request path, multi-agent desk, Strands loop.

Challenges we ran into

Bedrock invoke is unauthorized on this AWS account. Models list. Converse returns ValidationException: Operation not allowed. get-foundation-model-availability returns authorizationStatus: NOT_AUTHORIZED for Nova, Llama, Mistral, and Claude, in several Regions, including as root. That is an account-level hold, not a missing IAM tick. Nova is not a Marketplace subscribe model. A Bedrock API key does not bypass it. A Support case was opened.

I did not fake a Bedrock trace. The Strands loop is real. The published demo uses scripted tools so judges see observe → plan → act without waiting on a blocked API.

Local LLMs (LM Studio, Ollama/MLX) run the same tools. Reasoning models were too slow for a five-minute video. Instruct 3B is usable; the desk is not a token window.

Two AWS identities (classic account vs a new-experience project) overwrote the CLI profile. Credits sit on the classic account. We stopped creating extra projects.

Prompt-only guardrails fail. “Do not unlock” in the prompt was not enough. The hook has to cancel the tool.

What we learned

• Professional agents should take the pile, not the judgement. • Python as system of record is easier to evaluate than asking the model to “be careful.” • Strands is the agent. Bedrock is only the brain. If the brain is blocked, keep the tools. • GetFoundationModelAvailability tells the truth faster than the playground. • A live demo that lies about its backend is worse than a scripted demo that names its tools.

What's next

When AWS sets authorizationStatus to AUTHORIZED: MARKMAP_DESK_MODEL=bedrock, App Runner instance role, EventBridge POST /api/desk/sweep. AgentCore is optional later, not a second product.

Until then, the desk already runs.

Built With

• agents
• amazon
• amazon-web-services
• app
• bedrock
• boto3
• css
• docker
• eventbridge
• fastapi
• html
• iam
• javascript
• json
• lm
• ocr
• ollama
• pytest
• python
• runner
• sdk
• strands
• studio
• tesseract
• uvicorn