AI/ML Products

Multi-Agent HR Intelligence System (Vibe Coding)

July 2025- Present

Market Need & Problem Solving

The "Predictability" Gap: Traditional hiring often lacks data-driven predictability. This project solves this by quantifying job dimensions (Team, Business, Technical) to ensure a better fit.
Resumes are Static: The market is moving toward "Vertical Assessment." This project automates the creation of those assessment benchmarks, moving beyond what is written on a standard CV.
The Learning Curve Challenge: Companies struggle to measure how quickly a hire will provide value. The Dimension Difficulty Analyst agent specifically identifies the "difficulty level" of roles to help HR teams plan for onboarding and learning curves.
Efficiency in Scaling: Rapidly growing companies (like the 1-500 employee startups targeted) cannot manually draft high-quality, AI-ready job descriptions at scale. This tool reduces that manual overhead.

Technical Deep Dive: Agentic Flow Engineering

Stateful Orchestration (LangGraph): Moves beyond linear chains by using a StateGraph architecture to manage data persistence across nodes and enable conditional error-handling routes.
Precision Prompting: Employs Role-Based Personas and Few-Shot learning within structured templates to transform raw business inputs into quantified hiring dimensions.
LLM Optimization (OpenAI): Utilizes gpt-4o-mini with task-specific hyperparameter tuning—prioritizing creative reasoning (0.7 temp) for drafting and deterministic logic (0.1 temp) for scoring.
Validated Integration (FastAPI/LangChain): Uses a FastAPI backend to enforce Pydantic data validation before executing complex asynchronous logic via the LangChain Expression Language (LCEL).
Cloud Deployment (Railway/Gradio): Deploys the engine as a scalable SaaS API on Railway, interfaced through a Gradio dashboard that features real-time visualization of the agentic workflow.

Strategic Impact Tags

Vertical Assessment: Moving beyond static resumes to quantify Technical, Business, and Team dimensions.
Hiring Predictability: Solving the "Predictability Gap" through automated, data-driven candidate scoring.
Efficiency Metrics: Reducing Time-to-Hire (TTH) and Time-to-Productivity (TTP) by automating labor-intensive HR workflows.
AI Readiness Scoring: Measuring domain-specific AI competency to future-proof talent acquisition.

CarJudo: The Intelligence Layer for Automotive Assets

Jan 2026- Present

CarJudo is an AI-driven "Decision Intelligence" engine designed to de-risk the $1 Trillion U.S. used car market. While traditional aggregators focus on inventory and historical logs, CarJudo operates as a buy-side fiduciary, transforming fragmented data into predictive risk profiles. By narrowing a user's choice to the most mechanically sound and financially viable models, it generates high-intent, mature leads for the broader automotive ecosystem.

Metric Value Delivered

Hidden Liability Gap Converts vague mechanical anxiety into a quantified financial signal by identifying an average of $3,000–$5,000 in unpriced annual maintenance exposure.
Lead Maturity Rate Increases conversion for partners by delivering customers who have already cleared the 15-hour "Synthesis Burden" and are ready to transact.
Research Velocity Replaces fragmented manual searches with a 60-second 1–100 Confidence Score, accelerating the decision-to-purchase timeline.
Predictive Cost Transparency Provides a forward-looking view of ownership, ensuring that the annual cost of upkeep is factored into the initial purchase decision.

Data Model: BabsonAI Labs Operational Data Architecture

Sept- Dec 2025

Objective: Designed and implemented a relational database for a budding AI start-up to eliminate data silos, prevent redundant entries, and enable automated operational reporting.

Data Modeling: Developed a comprehensive Entity-Relationship Diagram (ERD) covering 7 core entities (Members, Projects, Clients, Tasks, Trackers, and Expenses) with defined cardinalities and referential integrity (Cascading/Non-Cascading).
Technical Implementation: Normalized and migrated mock datasets (100+ task entries, 500+ expense records) into DBeaver using SQL.
Business Intelligence: Authored complex SQL queries (Joins, Aggregations, Window Functions, and CASE statements) to transform raw operational data into actionable management insights.

Machine Learning Model: Credit Risk & Financial Distress Prediction

Jan- May 2025

Objective: Developed a predictive framework to identify at-risk borrowers likely to experience a financial crisis within two years, specifically aiming to balance high sensitivity (detecting defaults) with high specificity (minimizing false alarms for stable customers).

Data Engineering & Preprocessing

Dataset: 150,000 observations with 12 variables (e.g., Debt Ratio, Monthly Income, Credit Utilization).
Handling Missing Data: Used Predictive Imputation for 30,000 rows (Monthly Income and Dependents) to preserve data integrity over simple mean/mode substitution.
Feature Selection: Conducted correlation analysis to remove near-perfectly correlated variables (0.98–0.99), mitigating multicollinearity.
Class Imbalance Strategy: Addressed a severe 90/10 class split using Inverse Frequency Weighting, which amplified the influence of the minority "defaulter" class during training.