submission_metadata.yaml

# Redrob Hackathon — Submission Metadata

# ============================================================================
# Team identity
# ============================================================================
team_name: "CodeRed"

primary_contact:
  name: "J. Mohit Sai"
  email: "mohitjonnadula16@gmail.com"
  phone: "+91-9494366318"

team_members:
  - name: "J. Mohit Sai"
    email: "mohitjonnadula16@gmail.com"
    role: "Team Lead & Lead ML Engineer"
    contribution: "Co-architected the entire two-phase ranking pipeline and 4-pillar intent-aware scoring framework (Impact 42%, Ownership 33%, Search/JD Fit 15%, Behaviour 10%). Designed the NDCG@10-optimized weight allocation, additive boost and multiplicative penalty safety layers, and the evidence-graph reasoning engine with anti-hallucination safeguards. Built the V6.1 differentiator features including Tier-5 signature boost, behavioral twin penalty, pre-LLM x ownership interaction, and salary compatibility scoring. Led honeypot detection design (5-check system, ~80 traps) and developed the adversarial test suite for robustness validation."

  - name: "J. Mohan Bharath Chandra"
    email: "bharathjonnala123@gmail.com"
    role: "Lead ML Engineer"
    contribution: "Co-architected the entire two-phase ranking pipeline and 4-pillar intent-aware scoring framework. Designed and implemented the complete 55+ feature extraction engine across 13 groups (G1-G13) covering JD fit, impact evidence, ownership signals, production scale, career trajectory, retrieval depth, behavioural signals, and resume quality. Built all V7 new features including assessment signal, endorsement signal, education tier, and cross-validation. Implemented the deterministic tiebreak system with 14 sub-signals, career narrative analysis, and the LangChain-only penalty and closed-source isolation penalty modules."

  - name: "J. Venkatesh"
    email: "venkyismjamalapurapu09@gmail.com"
    role: "ML Engineer"
    contribution: "Implemented the TF-IDF + SVD (50-dim) deterministic semantic embedding pipeline with query expansion for synonym coverage. Built the adaptive Reciprocal Rank Fusion (RRF) retrieval layer and Parquet-based data handoff between precompute and rank phases. Optimized end-to-end pipeline runtime to ~200 seconds for 100K candidates with ~1.3 GB peak RAM on CPU-only infrastructure."

  - name: "M. Vineetha"
    email: "2300032887cseh1@gmail.com"
    role: "Data Engineer"
    contribution: "Implemented the two-phase pipeline orchestration (precompute.py and rank.py) with clean separation of concerns. Built the input parsing, JSONL loading, and submission CSV output modules. Handled data validation, candidate deduplication logic, and integration testing across the full pipeline. Managed dependency management and sandbox compatibility testing for reproducible execution."

# ============================================================================
# Code and reproducibility
# ============================================================================
github_repo: "https://github.com/Mohitsai3579/Redrob-hackathon"

sandbox_link: "https://huggingface.co/spaces/Mohit0708/Redrob-hackathon"

reproduce_command: "python rank.py --candidates ./candidates.jsonl --out ./submission.csv"

# ============================================================================
# Compute environment
# ============================================================================
compute:
  platform: "Windows"
  cpu_cores: 8
  ram_gb: 16
  python_version: "3.11.4"
  os: "Windows 11"
  uses_gpu_for_inference: false
  has_network_during_ranking: false
  pre_computation_required: true
  pre_computation_time_minutes: 4

# ============================================================================
# AI tools declaration
# ============================================================================
ai_tools_used:
  - "Claude"

ai_usage_summary: |
  Used Claude for architecture discussion and code review. No candidate data
  was fed to any LLM. All feature extraction and scoring is deterministic
  rule-based logic with no API calls.

ai_tools_notes: |
  All scoring weights, feature definitions, safety thresholds, and architectural
  decisions were made by the author. AI assistance was limited to code review
  and documentation drafting.

# ============================================================================
# Approach summary
# ============================================================================
methodology_summary: |
  Two-phase CPU-only pipeline. Phase 1 (precompute, ~196s) extracts ~55 JD-driven
  features per candidate across 13 groups — JD fit, impact/ownership, production
  and scale, experience and career, retrieval/evaluation, behavioural, resume
  quality, safety, location, career narrative, interaction features, key
  differentiators, and platform signals — then writes a typed Parquet feature
  store (~1.3 GB). Phase 2 (rank, ~4s) scores all 100K rows with a vectorized
  elite composite: impact 42%, ownership 33%, search/JD-fit 15%, behaviour 10%.
  Weights are modulated by hiring intent parsed directly from the JD (ownership
  expectation, team context, shipping culture, depth requirement).

  Safety stack: six multiplicative penalties ensure traps cannot reach the top 100.
  Honeypots are killed at 0.01×. JD disqualifiers (pure research, consulting-only,
  CV/speech primary, 18-month production gap) fire at 0.05×. Behavioral twins —
  strong on paper, recruiter_response_rate below 0.15 — receive a hard 0.0× filter.
  LangChain-only experience without pre-LLM ML depth is penalised at 0.55×.
  Closed-source isolation (5+ years, no external validation) at 0.60×. Title floor
  caps non-engineering profiles at 0.05–0.20×.

  Additive boosts reward deep specialists the keyword layer would otherwise miss:
  Tier-5 signature (+0.065), pre-LLM experience × ownership interaction (+0.045),
  impact × ownership combo (+0.035), and cross-validation across five quality
  dimensions simultaneously strong (+0.025).

  Keyword stuffing is blocked at the source: skills[] is excluded from the TF-IDF
  matching text. Only career_history descriptions, summary, and headline form the
  unified text blob, so a candidate who lists "RAG, Pinecone, FAISS" in skills
  but has no career context for any of them receives no skill-coverage credit.
  TF-IDF+SVD with query expansion provides deterministic, CPU-compliant semantic
  matching without bundling or downloading model weights.

  Reasoning is generated via an evidence graph: every technical keyword cited in
  the reasoning column is verified against the candidate's career text before
  emission, making hallucination structurally impossible. Template selection is
  hash-based (hashlib.md5 on candidate_id) for determinism and variation.

  Output is fully deterministic: REFERENCE_DATE is pinned, random_state=0 on SVD,
  no random module used anywhere. Running rank.py twice on the same machine
  produces byte-identical CSV output.

# ============================================================================
# Declarations
# ============================================================================
declarations:
  read_submission_spec: true
  code_is_original_work: true
  no_collusion: true
  honeypot_check_done: true
  reproduction_tested: true