Sushant
Kulkarni

Overview & Purpose

CricketIQ answers four core analytical questions: What is happening in a match? Why is it happening? What could happen next? What tactical options are available? Rather than acting solely as a live scorecard viewer, it models match dynamics through historical ball-by-ball datasets, predictive ML curves, counterfactual simulation, and tactical intelligence.

Key Engineering Features

• Match Intelligence: Complete match overview, dynamic scorecards, innings breakdowns, ball-by-ball delivery analytics, batting/bowling statistics, and continuous win-probability progression.
• Historical Win-Probability Model: Evaluates match states using runs required, balls remaining, wickets in hand, required run rate, and phase progression. (Evaluated historical deployment run documented in README: 364,406 rows, 84.38% accuracy, 0.10976 Brier score on evaluated historical data).
• Turning-Point Detection: Algorithms isolate momentum pivot points based on high-leverage deliveries, wicket clusters, and boundary streaks.
• Decision Replay: Counterfactual "what-if" analysis comparing real-world vs alternative strategies, estimating expected runs and win-probability deltas.
• Scenario Simulator: Interactive state modification (current score, overs, wickets lost, target) outputting projected totals, win probability, required run rates, and risk categorizations.
• Matchup & Player Intelligence: Granular batter-vs-bowler head-to-head metrics (strike rate, dismissals, dot-ball %, boundary %) and phase-wise splits (Powerplay, Middle, Death overs).
• Strategy Engine & Lab: Tactical recommendations synthesizing match situation, scenario simulation, and bowling/batting approaches (Aggressive Attack, Strike Rotation, Controlled Acceleration).
• Role-Based Authentication: JWT bearer authentication supporting FAN, ANALYST, COACH, and ADMIN tiers with an admin verification workflow.

System Architecture

External Cricket API ──► API Adapter ──► FastAPI Backend ──► PostgreSQL (Relational Master)
                                              │          ──► MongoDB (AI Reports)
                                              │          ──► Redis + Celery (Async Jobs)
                                              ▼
                                      React Dashboard (Vite + Recharts)
            

PostgreSQL: Relational master entities including Users, Teams, Players, Venues, Matches, Innings, Deliveries, and Decision Replays.
MongoDB: Schema-flexible document storage for generated analytical match reports.
Redis & Celery: Background task queues handling compute-intensive simulations and data processing.

Tech Stack & Infrastructure

Backend: Python, FastAPI, Pydantic, SQLAlchemy, Alembic, JWT, Celery

Frontend: React, Vite, React Router, Recharts, Framer Motion, Lucide React

Data & ML: PostgreSQL, MongoDB, Redis, Pandas, NumPy, scikit-learn

Deployment: AWS EC2, AWS RDS PostgreSQL, Docker Compose, Nginx, Let's Encrypt SSL, Elastic IP, DuckDNS, GitHub Actions

Problem Solved

Pharmaceutical catalogs suffer from inconsistent salt naming, disparate formulation syntax, varied packaging sizes, and complex multi-salt combinations. GenericRx normalizes unstructured drug compositions into canonical equivalents, enabling exact unit-price comparisons between expensive branded medicines and low-cost generics.

Key Engineering Features

• Fixed-Dose Combination Normalization: Splits compound salts across +, /, &, and and delimiters, strips conjugate variants, extracts numeric dosage strengths, and identifies release mechanisms (SR, PR, ER, XR, CR).
• Canonical Ingredient Sorting: Equivalent active ingredients are normalized and sorted into deterministic keys regardless of catalogue naming order.
• Price-per-Unit Standardization: Solves packaging skew by computing Price per unit = MRP / Pack Size, ensuring standardized fair comparisons.
• Indexed Fuzzy Search: PostgreSQL pg_trgm with GIN indexes powers prefix and typo-tolerant search, achieving sub-15ms response times.
• Prescription Savings Calculator: Multi-medicine regimen calculator projecting monthly and annual out-of-pocket savings.
• Asynchronous Scrapy Pipeline: Web crawler architecture coupled with Pandas, NumPy, and Regex cleaning modules processing a 3,000-medicine deployment catalog.

Data Pipeline Architecture

Medicine Sources ──► Scrapy Ingestion ──► Pandas Normalization Pipeline
                                                  │
                                                  ▼
                        PostgreSQL 16 (Master/GIN) + MongoDB 7 (Snapshots)
                                                  │
                                                  ▼
                              FastAPI Service ──► Next.js 14 Dashboard
            

PostgreSQL 16: Stores relational master catalogs, active salts, FK links, normalized unit prices, and GIN trigram indexes.
MongoDB 7: Stores raw scraped JSON snapshots, manufacturer notes, and raw archival documents.

Tech Stack & Infrastructure

Backend: Python, FastAPI, Pydantic, SQLAlchemy 2, Psycopg 3

Frontend: Next.js 14, React 18, Tailwind CSS, Lucide Icons

Data Pipeline: Scrapy, Pandas, NumPy, Python Regex

Infrastructure: AWS EC2, Docker Compose (Nginx, Next.js, FastAPI, PostgreSQL, MongoDB), Let's Encrypt SSL

Goal & Methodology

Evaluates resume fit against job descriptions using an explainable multi-tier matching architecture. Rather than relying on rigid keyword counters, it fuses lexical precision, dense semantic embeddings, and taxonomy-backed skill extraction.

Key Engineering Features

• TF-IDF Lexical Matching: Sublinear term-frequency scaling, word n-grams, and cosine similarity to measure direct technical terminology overlap.
• Dense Semantic Similarity: Sentence Transformers (all-MiniLM-L6-v2) with dense embeddings and cosine similarity to identify contextual equivalence across differing phrasings.
• Taxonomy-Driven Skill Extraction: spaCy PhraseMatcher mapped against a curated technical-skill ontology (Python, FastAPI, Docker, SQL, AWS, React, etc.).
• Composite Fit Scoring: Fuses lexical score, semantic embedding similarity, and technical skill coverage into a weighted composite metric.
• ATS Keyword Gap Analysis: Isolates high-leverage terminology present in the target job description but missing from the resume.
• Resume Tailoring Sandbox: In-browser workspace allowing candidate resume edits with instant score re-evaluation.
• Batch Processing & Export: Multi-resume candidate ranking against a single job description with structured HTML and JSON reporting.

Matching Pipeline

Resume + Job Description ──► Text Preprocessing
                                    │
           ┌────────────────────────┼────────────────────────┐
           ▼                        ▼                        ▼
   TF-IDF (Lexical)     Embedding (all-MiniLM-L6-v2)  spaCy PhraseMatcher
           │                        │                        │
           └────────────────────────┼────────────────────────┘
                                    ▼
                         Keyword Gap Analysis
                                    ▼
                    Composite Fit Score + ATS Guidance
            

Tech Stack

Core: Python, Streamlit UI, scikit-learn, sentence-transformers (all-MiniLM-L6-v2), spaCy, NumPy, HTML/JSON Reporting

Overview & Purpose

An automated audit platform analyzing review corpora and pricing patterns across Amazon, Flipkart, and Google Maps to detect fraudulent review campaigns, rating polarization, and e-commerce dark patterns.

Key Engineering Features

• Asynchronous Scraper Fleet: Headless Selenium drivers coordinated by Celery workers scraping reviews, star ratings, reviewer profile histories, and historical prices.
• Live Progress WebSockets: Django Channels, Daphne, and Redis Pub/Sub streaming real-time scraping and processing logs directly to the React dashboard.
• Review Velocity Anomalies: Statistical spike detection identifying suspicious volume bursts over constrained temporal windows.
• Semantic Duplicate Detection: Sentence Transformers (all-MiniLM-L6-v2) with PyTorch and cosine similarity to isolate paraphrased bot reviews.
• Aspect-Based Sentiment Analysis: Domain-tailored extraction using spaCy and NLTK VADER across retail (Battery, Camera, Build), dining (Food, Hygiene), and services.
• Dark Pattern Engine: Flags artificial pre-sale price markups, recurring artificial scarcity prompts, and single-review account clusters.
• True Trust Score: Composite trust score balancing anomaly frequencies, sentiment distribution, and reviewer credibility.
• Chrome Extension: Manifest V3 browser extension enabling on-page auditing of e-commerce listings.

System Architecture

Amazon / Flipkart / Maps ──► Selenium Workers ──► Celery Task Queue
                                                        │
                                                        ▼
                                       NLP Duplicate & Aspect Pipelines
                                                        │
                                                        ▼
  React Dashboard ◄── WebSockets (Daphne) ◄── Django REST + PostgreSQL
         ▲
         └── Chrome Extension (Manifest V3)
            

Tech Stack & Cloud

Backend: Python 3.12, Django 6, Django REST Framework, Daphne, Celery, Redis, Django Channels

NLP & ML: Pandas, NumPy, scikit-learn, Sentence Transformers, PyTorch, spaCy, NLTK

Ingestion: Selenium, BeautifulSoup4, WebDriver Manager

Frontend & Extension: React, Vite, Tailwind CSS, Recharts, Chrome Manifest V3

Cloud: Vercel (Frontend), Render (Backend), Upstash Redis, PostgreSQL

Engineering Philosophy & Scope

A probability-of-default modeling pipeline built around rigorous data hygiene: point-in-time SQL correctness, feature stability testing, scorecard calibration, and automated mutation testing to guarantee zero data leakage.

Key Engineering Features

• SQL Feature Engineering: CTEs and window functions in SQLite bounding peer and cohort aggregates strictly to earlier vintages, eliminating future information leakage.
• Point-in-Time Correctness Controls: Automated invariant tests verifying target leakage, future-vintage leakage, and peer aggregate boundaries.
• WOE Binning & IV Screening: Quantile binning, Weight of Evidence (WOE), and Information Value (IV) filtering to eliminate unstable or noisy variables.
• Feature Stability via PSI: Population Stability Index (PSI) tracking to catch distribution drift within the training sample and across validation cohorts.
• Explainable Logistic Scorecard: Points-scaled scorecard mapping probabilities to points with automated adverse-action reason codes.
• Mutation Testing Harness: A dedicated testing harness that deliberately introduces SQL leakage into the feature pipeline to verify that test suites catch it.
• Out-of-Time Validation: Chronological vintage holdouts rather than purely random train/test splits.
• Statistical Metrics: Implements Gini, Kolmogorov-Smirnov (KS), Brier Score, calibration curves, and bad-rate curves. (Note: Synthetic fixture numbers describe the generator's behavior, not a live loan book).

Pipeline Workflow

Raw Loan Data ──► SQL Feature Engineering (CTEs + Window Functions)
                          │
                          ▼
            Point-in-Time Leakage Invariant Tests
                          │
                          ▼
             WOE Binning + IV Screening + PSI
                          │
                          ▼
         Logistic Regression Scorecard + Adverse Action Codes
                          │
                          ▼
             FastAPI Serving Endpoint in Docker
            

Tech Stack

Core: Python, Pandas, NumPy, SQL, SQLite, scikit-learn (challenger models), MLflow, FastAPI, Docker, pytest

Overview & Architecture

A full-stack SPA providing polymorphic document management across standard Markdown notes, interactive checklists, and syntax-highlighted code snippets within a single MongoDB collection.

Key Engineering Features

• Polymorphic Document Schema: Three document variants (standard, checklist, code) sharing common metadata (title, tags, pin/archive state, timestamps) alongside type-specific fields.
• Weighted Full-Text Search: MongoDB text index queries with prioritized relevance weighting across titles, Markdown content, code explanations, and checklist items.
• Live Markdown Preview: GitHub-Flavored Markdown (GFM) editor supporting tables, task lists, and syntax blocks via react-markdown and remark-gfm.
• Interactive Checklists: Item addition, dynamic state toggling, completion percentage tracking, and atomic persistence.
• Code Snippet Storage: Language metadata tagging, inline explanation notes, copy-to-clipboard actions, and snippet categorization.
• Workspace Analytics: MongoDB aggregation pipelines calculating total notes, active/archived ratios, type breakdowns, and tag distributions.
• Layered Backend Architecture: Strict separation of concerns: API Router ➔ Pydantic Validation ➔ Repository Layer ➔ Motor Async MongoDB ➔ Serialized Response.

System Architecture

React 18 SPA (Vite + Tailwind) ──► Nginx Reverse Proxy
                                          │
                                          ▼
                      FastAPI (Router ➔ Pydantic Validation)
                                          │
                                          ▼
                         Repository Layer (Motor Async Driver)
                                          │
                                          ▼
                                      MongoDB 7
            

Tech Stack & Infrastructure

Backend: Python 3.11, FastAPI, Pydantic v2, Motor (Async MongoDB), Uvicorn

Frontend: React 18, Vite, Tailwind CSS v4, Lucide React, Axios, React Markdown, Remark GFM

Database & DevOps: MongoDB 7, Docker, Docker Compose, Nginx, Mongo Express