Experience the polished showcase and judge-facing presentation:
https://talentlens-inky.vercel.app/submission_demo.html
Explore the complete TalentLens workspace with explainability, analytics, comparisons, and candidate insights:
https://talentlens-inky.vercel.app/talentlens_x.html
Access the main entry point to TalentLens:
https://talentlens-inky.vercel.app/
Download the final ranked shortlist of 100 candidates:
- submission.csv (CSV format)
- submission.xlsx (Excel format)
GitHub Repository:
https://github.com/ruthvikgoud16/talentlens
Personal GitHub:
https://github.com/ruthvikgoud16
LinkedIn:
https://linkedin.com/in/ruthvikgoud
| Resource | URL |
|---|---|
| Homepage | https://talentlens-inky.vercel.app/ |
| Submission Demo | https://talentlens-inky.vercel.app/submission_demo.html |
| Interactive Dashboard | https://talentlens-inky.vercel.app/talentlens_x.html |
| GitHub Repository | https://github.com/ruthvikgoud16/talentlens |
| Developer Profile | https://github.com/ruthvikgoud16 |
https://linkedin.com/in/ruthvikgoud |
|
| Download Shortlist (CSV) | submission.csv |
| Download Shortlist (Excel) | submission.xlsx |
Recruiting at scale is a dynamic resource-allocation problem that is traditionally handled with static keyword-matching filters (ATS). This approach is highly fragile and prone to:
- Spelling and Conceptual Synonyms Failure: Missing a candidate with
"sentence transformers"or"llms"when the search is for"sentence-transformers"or"llm". - Generic descriptions / Buzzword stuffing: Candidates claiming expert skills without any backing evidence or projects in their career descriptions.
- Honeypot Contradictions: Candidates declaring senior-level titles or AI expertise when their technical career timeline shows only junior roles, unrelated domain tasks (e.g., pure manual testing), or impossible metrics.
- Demographic Biases: Biases introduced by name, location, gender pronouns, and age proxies that distort candidate merit.
TalentLens is a production-grade, de-biased, multi-signal candidate discovery and ranking platform. It processes 100,000 candidates in under 3 minutes, using local, deterministic execution vectors to evaluate candidates along four key dimensions:
- Demographic De-Biasing & Normalization: Strips pronouns, ages, locations, and names before candidate scoring to prevent cognitive biases.
- Spelling-Aware Skill Extraction: Pre-normalizes skills and matches synonyms to avoid false negatives.
- CEILING WEIGHTS & Core Verification: Assesses candidates against a realistic skill ceiling (
SKILL_CEILING_WEIGHT = 13.0points) instead of demanding matches across all 25+ JD keywords, which avoids score compression. - Honeypot Contradiction Checking: Flag and penalize impossible claim-to-evidence timelines (like CV/Robotics specialists applying for NLP search roles, or junior developers claiming senior engineering titles).
- Recruitability & Availability Metrics: Incorporates behavioral and logistical signals (notice period, interview completion rates, GitHub activity) to compute hiring feasibility.
+-------------------------------------------------+
| Input Job Description |
+-------------------------------------------------+
|
v
+-------------------------------------------------+
| Module 1: Demographic De-Biasing Layer |
| (Names, pronouns, locations redacted) |
+-------------------------------------------------+
|
v
+-------------------------------------------------+
| Module 2-3: Semantic Matching (Ceiling) |
| (TF-IDF, synonym-matching skill extraction) |
+-------------------------------------------------+
|
v
+-------------------------------------------------+
| Module 4-5: Honeypots & Contradiction Audits |
| (Timeline continuity, consulting ratios) |
+-------------------------------------------------+
|
v
+-------------------------------------------------+
| Module 6: Deterministic Score Aggregation |
| (0.35x Career + 0.25x Skills + 0.25x Recruit |
| + 0.15x Semantic) |
+-------------------------------------------------+
|
v
+-------------------------------------------------+
| Module 7-9: Recruiter Dashboard & Analytics |
| (Top 100 csv + interactive comparisons) |
+-------------------------------------------------+
The search pipeline filters candidates systematically at increasing levels of resolution to ensure high latency throughput:
- Pipeline Ingestion: 100,000 candidate profiles parsed and anonymized (latency: ~2.5 min).
- Semantic Filter: Top 1,000 candidates retrieved via token cosine similarity.
- Aggregated Scoring: Deterministic calculation on career, skills, and recruitability (latency: <500ms).
- Shortlist Generation: Top 100 candidates written to
submission.csvwith recruiter-facing CoT justifications.
- De-Biasing Pronoun Mapping: Automatically converts he/she/his/her to they/their, redacts names to
[NAME_REDACTED], and masks locations. - Skill Spelling Normalization: Resolves spacing, casing, and hyphens (e.g.
FAISSvsfaiss,sentence transformersvssentence-transformers). - Core Domain Hardening: Penalizes backgrounds in computer vision, speech recognition, and robotics (e.g.,
-45points career penalty) to strictly align with the Information Retrieval/Search specification. - Notice Period and Logistics weighting: Prioritizes immediate joiners (≤30 days notice) while adjusting recruitability score downwards for passive candidates or long notices.
- CoT AI Reasoning Trail: Local CLI engine builds recruiter narrative summaries, explaining exactly how each score was derived based on verifiable timeline evidence.
TalentLens provides full transparency for every ranking. Re-ranked candidate cards expose a detailed Evidence Panel that lists the exact matching sentences and projects extracted from their career descriptions. In addition, the comparison dashboard allows recruiters to evaluate the top two candidates side-by-side:
- Radar Chart Overlay: Multi-dimensional visualization of candidate competencies (Semantic, Truth, FutureFit, Experience, GitHub, Recruitability).
- Delta Breakdowns: Granular score comparisons detailing exactly why Rank #1 beats Rank #2 (e.g. scale advantage: 50M users vs 2M queries).
When launching talentlens_x.html, you can trigger the Autoplay Demo Mode that showcases the entire candidate discovery pipeline:
- JD Parsing: Paste a Job Description and watch the NLP engine extract structural keywords.
- Semantic Discovery: Watch the skeleton loaders shimmer as candidate records are parsed and mapped to the TF-IDF space.
- Shortlist Rankings: Display the candidate list with custom gold, silver, and bronze card outlines for the top 3.
- Compare Candidates: Select the top candidates to inspect their evidence panel, timeline badge highlights, and radar competence overlay.
- CSV Download: Export a validator-compliant
submission.csv.
- Front-End Fallbacks: Both dashboard apps are fully client-side compatible. If Groq/Gemini API calls fail, the interface switches to a client-side parser fallback, enabling complete local interactive walkthroughs.
- No Network at Inference:
app.pycomputes all scores locally on your machine without relying on active network calls, ensuring fast, stable, and secure evaluations.
- Backend Logic: Python 3.9+, Pandas, NumPy, Scikit-Learn (TF-IDF Vectorization)
- Frontend Dashboard: Vanilla HTML5, Vanilla CSS3 (Glassmorphic variables design system), JavaScript (ES6, Native SVG Canvas charts)
talentlens/
│
├── assets/
│ ├── screenshots/
│ │ ├── dashboard_view.png # Interactive app landing screenshot
│ │ ├── workflow_stepper.png # Live stepper animation screenshot
│ │ ├── candidate_comparison.png # Side-by-side Rank 1 vs 2 screenshot
│ │ ├── top_podium.png # Top 3 candidate podium screenshot
│ │ ├── evidence_panel.png # Truth Engine checked evidence screenshot
│ │ ├── analytics_charts.png # SVG charts screenshot
│ │ └── talentlens_architecture.png # System architecture diagram
│
├── app.py # Local Candidate Ranking Engine
├── test_app.py # Integration testing suite
├── talentlens_x.html # Interactive Dashboard
├── submission_demo.html # Static Submission Results Viewer
├── submission.csv # Target Ranked Top 100 CSV Output
├── submission_metadata.yaml # Manifest YAML file
├── job_description.txt # Ingested Senior AI Engineer spec
├── requirements.txt # Python packages list
├── LICENSE # MIT License
└── .gitignore # Ignored dataset/caches config
Ensure you have Python installed, then install packages:
pip install -r requirements.txtEnsure candidates.jsonl is placed in the project directory, then run:
python app.pyThis processes the dataset and regenerates submission.csv containing the ranked top 100 candidates with descriptive recruiter reasons.
To run unit assertions verifying the scoring logic against mock candidates (Perfect AI, CV Specialist, Consulting Developer):
python test_app.pyOpen your browser and navigate to either:
- talentlens_x.html to run the interactive dashboard.
- submission_demo.html to view the static candidate comparison view.
- Top Candidate (Rank #1):
CAND_0086022(Final Score:77.24). Senior Applied Scientist with strong embeddings, vector DB, and retrieval experience. 4+ embedding hits, GitHub activity score 75/100. - Monotonically Decreasing Scores: Candidate scores range from
77.24down to approx54for Rank #100. - Ties Handled Compliantly: Ties broken using ascending candidate ID as tiebreaker.
- Ruthvik Goud — Sole Developer & AI Architect
- GitHub Profile: @ruthvikgoud16
- LinkedIn: Ruthvik Goud Profile
- Email: bathiniruthvik370@gmail.com
- Phone: +91-9392796458
- Credits: Developed for the Redrob Hackathon. Evaluation templates, test runner specifications, and initial candidate data schema supplied by the Redrob team.
- License: Distributed under the open-source MIT License.