I design and build verifiable AI systems, autonomous simulation agents, and security-critical backend architectures. My work bridges probabilistic machine learning models with deterministic software guarantees—combining classical statistical learning, graph algorithms, and schema-constrained LLM reasoning.
- Research Interests: Hybrid ML-LLM Pipelines, Graph-Based Anomaly & Fraud Detection, Sequential Decision-Making Agents, Hybrid Information Retrieval (
BM25+ Dense Embeddings), and Multimodal Computer Vision. - Open-Source & Systems Engineering: Building modular
Python/PyTorch/FastAPIarchitectures, Model Context Protocol (MCP) servers, and developer tooling (such as Codecritter) at @Buildingcoolstuff. Selected as an Official Contributor for Open Source Connect India 2026. - Mathematical Foundations: Implementing core machine learning algorithms, linear algebra routines, and deep neural network architectures from scratch in
NumPyandPyTorch(deep-ml-journey & 300-Days-of-AI-ML).
1. Sentinel — Hybrid Graph-ML & Schema-Bound LLM Risk Engine
Isolation Forest Gradient Boosted Trees Graph Clustering Pydantic MCP Server FastAPI
- Ensemble & Graph Architecture: Engineered a multi-stage fraud-ring and abuse-spike detector combining an Isolation Forest + GBDT ensemble with graph-based fraud-ring clustering (built for Razorpay AI Buildathon 2026).
- Deterministic Safety Layer: Integrated a schema-bound LLM reasoning layer (
Pydantic-validated with deterministic fail-safe fallbacks) that engages exclusively on high-entropy, ambiguous edge cases—ensuring zero automated account freezes without human-in-the-loop audit trails. - Systems & Verification: Exposed core risk evaluation pipelines as verified Model Context Protocol (MCP) tools and validated system robustness against a 21-test adversarial suite.
2. Kaggriculture Agent — Autonomous Multi-Turn Simulation Agent
Sequential Decision-Making Heuristic Optimization Multi-Agent Simulation Python
- Stochastic Policy Design: Designing an autonomous agent operating over a 720-turn stochastic horizon for Kaggle's $50K simulation benchmark, optimizing multi-variable resource allocation, crop yield mechanics, and economic state transitions under uncertainty.
- Evaluation & Telemetry Harness: Built local diagnostic benchmarking harnesses and state-tracking telemetry validated directly against the official competition simulation engine to test policy updates and prevent regression.
3. Hybrid ATS Resume & Information Retrieval Engine — Lexical + Semantic Ranking Pipeline
BM25 / TF-IDF Dense Embeddings Supervised Classification FastAPI React Docker
- Two-Stage Ranking Architecture: Constructed a document-to-query matching pipeline combining sparse lexical retrieval (TF-IDF / BM25) with dense semantic vector embeddings, feeding into a calibrated shortlist classifier and a structured LLM diagnostic layer.
- Reproducible Deployment: Engineered as an end-to-end containerized microservice using FastAPI, React, and Docker for deterministic local and cloud inference.
4. ClassroomAI Monitor — Real-Time Multimodal Vision & Audio Telemetry
YOLOv8 MediaPipe 3D Pose OpenAI Whisper FastAPI Streams React
- Multimodal Fusion Pipeline: Built a low-latency behavioral telemetry system (🥇 1st Place Winner, BYTEHACK 2026) fusing YOLOv8 multi-person detection, MediaPipe 3D head-pose and kinematic gesture tracking, and Whisper real-time speech transcription over asynchronous FastAPI + React streams.
Rather than treating libraries as black boxes, I maintain open-source repositories dedicated to mathematical derivations, algorithmic complexity, and modular developer tools:
| Repository / Initiative | Scope & Technical Focus | Status / Milestone |
|---|---|---|
| 300-Days-of-AI-ML | End-to-end progression from statistical foundations and unsupervised clustering (K-Means, DBSCAN, Agglomerative) to PyTorch Deep Learning architectures built from scratch. |
Day 230 / 300 (77%) ███████████████████████░░░░░░░ |
| deep-ml-journey | Daily from-scratch implementations of linear algebra kernels, optimization routines, and ML algorithms in pure NumPy and PyTorch. |
Active Daily Repository |
| @Buildingcoolstuff | Engineering organization for open-source developer tools (including Codecritter), applied AI workflows, and reproducible backend systems. | Lead Maintainer |
| Open Source Connect India 2026 | Contributing to open-source Python backend infrastructure, automated security tooling, and reproducible ML workflows. | Selected Official Contributor |
| Competition / Benchmark | Distinction | Technical Deliverable / Domain |
|---|---|---|
| BYTEHACK 2026 (BCE Patna) | 🥇 1st Place Winner | Engineered ClassroomAI Monitor — real-time multimodal CV & audio telemetry pipeline. |
| RookieCTF 2026 (IIT Patna — Tatva) | 🥇 1st Place Winner | Scored 3,648 pts with team The Cypher Crew across Cryptography, Digital Forensics, OSINT, and Web Exploitation. |
| 29th National Youth Festival 2026 | 🥇 1st Place (Software) | Built CampusGigs — peer-to-peer micro-economy platform with LLM task structuring (Live Demo). |
| Adobe University Hackathon 2026 | 🎯 Top ~9% Nationwide | Cleared Round 1 — selected among 26,760 teams out of 301,082 registered teams nationwide. |
| Layer | Technologies & Tools |
|---|---|
| Languages | Python · C · Java · JavaScript (ES6+) · SQL · Bash |
| ML, Vision & Research | PyTorch · scikit-learn · NumPy · Pandas · YOLOv8 · MediaPipe · OpenCV · Whisper · BM25 / Embeddings |
| Backend, APIs & Infra | FastAPI · Docker · Linux (Ubuntu) · Git · GitHub Actions (CI/CD) · Model Context Protocol (MCP) · Pydantic · Supabase |
| Frontend & Interfaces | React · Next.js · Tailwind CSS · Vite |
| Security & Verification | Adversarial Testing · Applied Cryptography · Digital Forensics · OSINT |
"Build from first principles. Verify under adversarial conditions. Ship to production."