User sees
A calm coaching surface
A simple chat and a plan that updates when the user needs direction.
System handles
The reasoning underneath
Routing, memory, safety context, and normalized plan changes before the answer reaches the user.
Case Study / Early Prototype
AI OrchestrationSpecialist RoutingPlan Memory
01 / OVERVIEW
An AI gym coach
built around structure,
not louder advice.
MARC gives users calm, practical training guidance while specialist agents handle progression, recovery, pain signals, nutrition, and plan continuity underneath.
02 / SYSTEM PIPELINE
The system behind the simple chat
The user mostly sees a coach and a plan. That was intentional. Underneath it, MARC keeps track of context, routes the request when specialist reasoning is needed, and turns the response back into something usable in the gym.
User signal
Goal, context, constraints, and current state.
Coach router
Clarifies intent and routes to the right specialist.
Plan memory
Stores plan state, decisions, and context over time.
Normalized answer
Returns structured guidance for safe, practical action.
03 / PRODUCT IN ACTION
Practical guidance with a system behind it.
The useful part of MARC happens after the conversation. The plan updates, the context carries forward, and the next action becomes easier to follow.
Routing logic01
Specialist routing
Escalations happen only when recovery, pain, nutrition, or progression signals need deeper reasoning.
Plan memory02
Context Carries Forward
Training decisions stay connected across sessions instead of resetting every time the user sends a new message.
Usable Output03
Plan-shaped answers
The important output is a clearer plan, a safer adjustment, or the next action, not just a better paragraph.
Coach chat
The chat creates and updates the plan
The coach asks for context, builds a plan, and keeps the user moving toward the next action.
04 / SYSTEM DECISIONS
Keeping the interface simple meant making the system stricter.
MARC’s goal was always to feel straightforward: talk to the coach, get the plan updated, keep training. The harder work sat underneath that surface. The system needed to know when to route, what to remember, and how to return advice without making the user decode the machinery behind it.
Feature01
AI GUIDANCE
Generated from structured context, not generic coaching prompts.
Routing02
SPECIALIST ROUTING
Recovery, nutrition, pain, and progression questions can trigger deeper specialist reasoning.
CONTEXT03
PLAN MEMORY
Training decisions stay connected across sessions instead of starting from scratch every chat.
OUTPUT04
NORMALIZED OUTPUTS
Responses are shaped into plan changes, safety adjustments, or next actions.
05 / LEARNING LOG
The build showed me where AI coaching gets awkward.
MARC was less about making a chat interface and more about learning what has to happen around the chat. A coach needs memory, routing, restraint, and outputs that do something useful after the answer is generated.
~/marc/learning-log.md
READ ONLY
mrinal@portfolio:~$ cd marc
mrinal@portfolio:~/marc$ less learning-log.md
LOG 01
Interfaces need restraint
MARC worked best when the interface stayed quiet. The product had to make complex coaching logic feel simple enough to trust in the middle of a workout.
LOG 02
AI needs routing, not just prompting
Better prompts were not enough. The system needed to know when to involve specialist reasoning for recovery, pain, nutrition, or progression.
LOG 03
Plan needs memory
A useful coach cannot reset every conversation. Training decisions had to carry forward so each answer could respond to the current plan, not just the latest message.
LOG 04
Outputs need normalization
The strongest responses were shaped into clear plan updates, safety adjustments, or next actions. The value was in making the answer usable.
SYSTEM STATUS :BUILDING