Skip to content
HirthikBalajiPublic

About

Anokha Hacktide Event Repository Anokha 2026

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

Methodology

1. Core Contribution Equation (High-Level)

$$ \boxed{ \text{Contribution Score (CS)} = \Big( \sum_{i=1}^{N} W_i \cdot A_i \Big) \times Q \times C \times P - B } $$

This keeps it modular, explainable, and tunable.


2. Base Work Done (Activity Contribution)

$$ \text{Base Work} = \sum_{i=1}^{N} W_i \cdot A_i $$

Where:

  • ($A_i$) = Count or intensity of activity type i
  • ($W_i$) = Weight for that activity

Example Activity Variables

Activity ($A_i$) Description
($A_1$) Tasks completed (Jira)
($A_2$) Code commits
($A_3$) PRs merged
($A_4$) Meetings led
($A_5$) Design docs / proposals
($A_6$) Mentoring / reviews done

3. Quality Multiplier (Peer Review + Outcomes)

$$ Q = 1 + \alpha \cdot R + \beta \cdot S $$

Where:

  • (R) = Peer review score (0–1 or 1–5 normalized)
  • (S) = Stability / success rate (bug-free releases, acceptance rate)
  • ($\alpha, \beta$) = Tunable impact factors

📌 Ensures quality > quantity


4. Collaboration & Knowledge Sharing Factor

$$ C = 1 + \gamma \cdot K $$

Where:

  • (K) = Knowledge contribution index (helping teammates, answering Slack threads, docs, unblocking others)

5. Peer Recognition / Kudos Factor

$$ P = 1 + \delta \cdot \log(1 + U) $$

Where:

  • (U) = Peer kudos / endorsements
  • Log prevents popularity bias
  • ($\delta$) controls recognition influence

6. Blocker & Friction Penalty

$$ B = \lambda_1 \cdot D + \lambda_2 \cdot F + \lambda_3 \cdot R_w $$

Where:

  • ($D$) = Delays caused
  • ($F$) = Rework or rejected submissions
  • ($R_w$) = Repeated unresolved blockers
  • ($\lambda$) = penalty weights

🚨 Penalizes friction, not silence


7. Optional Impact Scaling (Business Value)

$$ \text{Final Score} = CS \times (1 + \theta \cdot I) $$

Where:

  • (I) = Business impact (revenue, customer impact, risk reduction)

8. Compact General Formula (Production-Friendly)

$$ \boxed{ \text{CS} = \Big( \sum W_i A_i \Big) \cdot (1 + \alpha R + \beta S) \cdot (1 + \gamma K) \cdot (1 + \delta \log(1 + U)) ;-; \sum \lambda_j P_j } $$


9. Why This Model Works Well (Important)

  • Role-agnostic (engineer, manager, designer)
  • Resistant to gaming
  • AI-friendly (features from Slack, GitHub, Jira)
  • Explainable to HR & leadership
  • Encourages collaboration over noise

About

Anokha Hacktide Event Repository Anokha 2026

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages