How TalentGraph thinks

Methodology & Explainability

A fully-explainable hybrid ranker. No hosted LLMs, no GPU, CPU-only, reproducible. Every score decomposes into the components below.

Scoring formula
FINAL = (BASE_FIT − penalties) × behavioral_mult × location_mult
honeypots → floored to ~0 (relevance tier 0)
BASE_FIT components (max 100)
Retrieval / Embeddings / Vector DB
22
Ranking / Search / RecSys
20
Evaluation Frameworks
12
Product-Company Engineering
14
Python + Production Depth
10
Career Trajectory & Seniority
10
Experience Band Fit
6
Open Source / GitHub
4
Corroborated Skills
2
Ranking pipeline
1
Parse & stream
Read 100K JSONL profiles in a memory-safe stream (~7s).
2
Extract evidence
Regex lexicons read career descriptions & summaries — retrieval, ranking, evaluation — weighted 2× over the skills list.
3
Detect honeypots
Profile-consistency checks floor impossible profiles (role > career, expert-with-0-months, broken chronology).
4
Score & penalize
Evidence-weighted BASE_FIT minus keyword-stuffing / services-only / research-only / off-target penalties.
5
Apply multipliers
Behavioral availability (0.55–1.15×) and location (0.90–1.05×) modifiers.
6
Rank & reason
Top 100 with monotonic scores, deterministic tie-break, and evidence-grounded reasoning.
Down-rank penalties
  • − Keyword stuffing (AI skills, no career evidence)
  • − Non-technical title with no engineering evidence
  • − Pure research-only, no production
  • − Services/consulting-only career
  • − Off-target CV / speech / robotics without NLP/IR
  • − Tutorial / framework-only flavour
  • − Over-experience drift (>13y) without recent production
Availability & location

A perfect-on-paper candidate who hasn't logged in for six months and answers 5% of recruiters is, for hiring purposes, not available. The behavioral multiplier (0.55–1.15×) folds in response rate, recency, notice period, and interview/offer reliability. A small location multiplier favours Pune/Noida and tier-1 India.