[Bug] : Duplicate Model Training Optimization: Implement Caching for Calibration Validation

[sohampirale]

Issue Description

Currently, the calibration validation endpoint (/api/session/calib_validation) performs the same expensive model training computation for both requests with from_ruxailab: false and from_ruxailab: true. This results in unnecessary duplicate computation since both requests use identical input parameters and should produce identical results.

---

Current Flow

1. Frontend sends first request with from_ruxailab: false
2. Backend calls gaze_tracker.predict() → performs full model training
3. Frontend sends second request with from_ruxailab: true
4. Backend calls gaze_tracker.predict() again → performs identical model training
5. Second request additionally sends results to Firebase webhook

---

Impact

• Performance: Double the computational cost for identical operations
• Resource Usage: Unnecessary CPU cycles and memory consumption

---

Proposed Solution

Implement a caching mechanism to store the results of gaze_tracker.predict() and reuse them for subsequent requests with identical parameters.

---

Caching Options Under Discussion

There are ongoing discussions about the best caching approach:

1. In-memory caching (LRU cache, file-based)
2. Shared memory solutions for multi-process environments
3. External caching services (Redis, etc.)

---

Related Discussions

Related Issue : #61
Related PR : #68

---

Request for Input

Appreciate the guidance from maintainers on:

1. Preferred caching mechanism for this use case
2. Configuration parameters (cache size, TTL, etc.)
3. Multi-process deployment considerations

---

Expected Benefits

• Significant reduction in computational overhead
• Faster response times for calibration validation
• Better resource utilization
• Improved user experience during calibration process

---

[sohampirale]

Working on this
Will create a PR once we finalize the caching approach based on the ongoing discussion.