File size: 4,095 Bytes
c754148 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 | team:
name: "Ctrl Coffee Repeat"
primary_contact_name: "Pranjal H Dohare"
primary_contact_email: "pranjaldohare8@gmail.com"
primary_contact_phone: "+919320480095"
github_repository_url: "https://github.com/Pranjal1342/Intelligent-Candidate-Discovery-Ranking-System"
sandbox_demo_url: "https://ctrl-coffee-repeat.streamlit.app/"
members:
- name: "Pranjal H Dohare"
role: "Lead Developer"
- name: "Priyanka Tiwari"
role: "Architecture and System Design"
submission:
version: "1.0.0"
timestamp: "2026-07-01"
output_file: "CTRL_COFFEE_REPEAT.csv"
system:
pipeline_type: "Offline-Indexed Lexical Retrieval + LightGBM LambdaRank"
hardware: "CPU-only, ≤16GB RAM"
runtime_seconds: 4
network_calls_during_ranking: 0
methodology_summary: |
This system uses a deterministic, CPU-only pipeline optimized for NDCG@10 and P@5.
Stage 1 (Retrieval): A precomputed NumPy CSR BM25 matrix (built offline, ~40 MB) is queried
at runtime in under 0.1 seconds via dual-pass: Pass A expands JD requirements using a
skill alias taxonomy (skill_aliases.json), Pass B targets production-context keywords
(deployed, scale, serving, latency). A rare-term safety net retrieves candidates with niche
skills (pinecone, lambdarank) that might otherwise be missed. This produces a ~8,500-candidate
Stage 1 pool in approximately 0.03 seconds.
Stage 2 (Features): A 22-feature schema-grounded matrix extracts signals from every candidate
record. Includes 5 adversarial detection functions: domain-category mismatch, synthetic template
detection, production signal log-compression, LangChain dabbler detection, and CV/speech
specialist detection. Stage 3 adds a consistency composite (c1×c2×c3×c4×c5) that zeros out
scores for timeline impossibilities, signup anomalies, salary inversions, assessment
contradictions, and engagement mismatches.
Stage 4 (Ranking): LightGBM with objective=lambdarank trains on relevance labels generated
via 2,500 pairwise LLM comparisons using Gemma3:4b-it-q4_K_M (running offline and locally
via Ollama — zero external API calls). This explicitly breaks circularity: the LLM judges
profiles organically without knowledge of the 22 features or BM25 scores, then Elo ratings
are converted to 0-3 relevance labels by quartile thresholding. Candidates with data integrity
violations are suppressed post-inference via a consistency multiplier
(final_score = raw_score × consistency_score).
Stage 5 (Reasoning): Deterministic grammar engine generates fact-grounded reasoning with
numeric regex audit (all cited numbers must exist in the candidate JSON), n-gram collision
avoidance (difflib.SequenceMatcher), and priority-ranked concern surfacing. Pre-submission
blocking audits enforce diversity (max 25% archetype concentration, max 30% employer
concentration) and honeypot detection (assert low_consistency_in_top100 < 10).
Model comparison evidence: the heuristic-trained model required a hand-coded suppression list
to keep non-technical profiles out of the top 100. The Gemma-trained model achieved 0 honeypot
leakage with no suppression list, and the two models show Spearman correlation of 0.001 on the
top-100 ranking — confirming the LLM labels are genuinely independent of the engineered features.
ai_tools_used:
- tool: "Google DeepMind Antigravity"
usage: "Code scaffolding, module structure, latency diagnostics, iterative debugging"
human_review: true
- tool: "Gemma3:4b-it-q4_K_M via Ollama (local, offline)"
usage: >
Offline pairwise candidate annotation: 2,500 comparisons on a stratified sample of
500 Stage 1 candidates to generate non-circular LightGBM training labels.
No candidate data transmitted to any external service. Runs in
experiments/pairwise_llm_check/annotate_and_retrain.py, entirely separate from
the ranking pipeline. Exempt from the 5-minute/zero-network ranking budget.
human_review: true
reference_date: "2026-01-01" |