I specialize in neural-symbolic reasoning and high-performance machine learning systems. My work combines systems programming (Rust) with applied ML research, focusing on mathematical reasoning, gradient boosting algorithms, and production ML deployment.

I build systems that are formally verified, performance-optimized, and production-ready.

AI researcher specializing in neural-symbolic reasoning and high-performance machine learning systems. Independently reproduced Google DeepMind's AlphaProof mathematical reasoning approach and authored a novel gradient boosting algorithm outperforming XGBoost by 18% on extreme class imbalance. Published researcher with production ML systems deployed at Fortune 500 companies achieving 100,000+ queries/second.

Core Expertise: ML algorithm design, Monte Carlo Tree Search, Rust systems programming, SIMD optimization, formal verification, and production ML deployment.

Mar 2025 – Present | Independent Research | Rust, 67k LOC

Reproduced Google DeepMind's 2024 AlphaProof approach: a hybrid neural-symbolic system combining Monte Carlo Tree Search with a Transformer policy network for automated mathematical reasoning.

Technical Achievements:

• Built 450+ formally verified transformation rules spanning IMO-level mathematics: algebra, calculus, trigonometry, number theory, inequalities, combinatorics, and polynomial manipulation
• Achieved 95.2% accuracy on single-step problems and 100% on multi-step problems across a 31-test benchmark suite
• Designed end-to-end training pipeline: auto-generated 17,000 synthetic mathematical problems and trained a Transformer for 50 epochs to guide symbolic rule selection
• Key innovation: Provides formally verified proof traces with complete step-by-step justification, eliminating hallucinated intermediate reasoning steps common in LLM-based systems

Architecture: Custom AST parser, MCTS with UCB selection, Transformer policy network (Candle), integrated numerical and symbolic verifier

Links: GitHub | License: Mozilla Public License 2.0

Jun 2025 – Present | Published Research | Rust + PyO3, 13K LOC

Novel gradient boosting algorithm fusing Shannon entropy with Newton–Raphson optimization, outperforming XGBoost by 17.9% PR-AUC and LightGBM by 10.4% on credit card fraud detection (0.2% minority class, 284K samples).

Technical Innovations:

• Extreme drift resilience: 1.8% degradation under covariate shift vs. XGBoost (31.8%) and LightGBM (42.5%)
• Systems-level optimizations: Zero-copy architecture (31.7 MB training overhead), cache-aware data structures (64-byte alignment), 8x loop unrolling for SIMD auto-vectorization, <5ms histogram construction
• 45-second training on 170K samples; supports binary classification, multi-class (One-vs-Rest with softmax), and regression
• Auto-tuning system that profiles dataset characteristics and derives optimal hyperparameters, eliminating manual tuning

Impact:

• Published: Zenodo DOI 10.5281/zenodo.17568991
• 4,300+ PyPI downloads — 60+ GitHub stars — Featured on Kaggle

Links: GitHub | PyPI | Install: pip install pkboost

Dec 2025 – Present | PKBoost AI Labs — Value Score Business Solutions | Rust + React, 6K LOC

Role: Lead systems engineer responsible for end-to-end architecture, performance optimization, and production deployment.

Built an ultra-high-performance document Q&A system achieving 100,000+ queries/second, <5ms vector search latency, and 300ms end-to-end response time including LLM inference.

Technical Achievements:

• 10–100x performance improvement over database-backed baselines (USearch in-memory HNSW: 5ms vs. PostgreSQL pgvector: 50ms for 10K-vector search)
• Production-grade security: JWT authentication, Argon2 password hashing, token-bucket rate limiting, SQL injection protection, CORS enforcement
• Multi-format ingestion pipeline (PDF, Excel, Word, text) with optional Tesseract OCR and semantic chunking using all-MiniLM-L6-v2 embeddings (384-dim)
• Fully async architecture using Tokio to handle 1,000+ concurrent connections with connection pooling and single-binary deployment (~50MB for 10K vectors)

Real Deployment: Deployed at a Fortune 500 company (Under NDA) supporting 1,000+ employees with <5ms semantic search across 10,000+ document chunks.

Tech Stack: Rust (Axum), Tokio, USearch, FastEmbed-rs, PostgreSQL, React + Vite, Groq API (Llama 3.3)

Closed Source project (Under NDA)

PKBoost AI Labs | Dec 2025 – Present | Mumbai, India

Founded independent AI/ML research lab focused on high-performance tabular ML, neural-symbolic reasoning systems, and production ML infrastructure.

• Research priorities: Concept drift adaptation, formal mathematical reasoning, SIMD-optimized inference, interpretable gradient boosting
• Built and maintained 3 major open-source projects (PKBoost, LEMMA, RAG) with 25K+ lines of production Rust code and active user communities

Value Score Business Solutions LLP | Jun 2025 – Present | Mumbai, India

• Architected and deployed agentic RAG workflows using n8n automation and open-source LLMs for document-based question answering
• Built production Rust RAG agent with USearch vector search—demoed to a Fortune 500 for employee HR assistance (1,000+ user capacity)
• Developed LLM-powered email personalization system with Groq/Grok validation and quality checks
• Evaluated Zoho Catalyst platform for ML model deployment and CRM integration

Artech Communications | Feb 2025 – Apr 2025 | Mumbai, India

• Configured high-availability hospital LAN with redundancy and failover
• Administered Linux/Windows servers with security hardening and validation
• Performed penetration testing and network security audits

Diploma in Computer Technology | 2022 – 2025
K.V.M Institute of Technology — Mumbai, India | CGPA: 8.1/10

• Reproduced cutting-edge AI research (Google DeepMind's AlphaProof, AlphaZero)
• Published novel ML algorithm with formal benchmarking and evaluation
• Mentored by Ash Vardanian (Founder, Unum Cloud; Creator of USearch, SimSIMD)

• Advanced Machine Learning: Gradient boosting internals, MCTS, Transformers
• Systems Programming: Cache optimization, SIMD vectorization, Rust concurrency
• Mathematics: Information theory, numerical optimization, linear algebra
• Formal Methods: Symbolic verification, proof systems, type theory

• Published researcher: Zenodo DOI 10.5281/zenodo.17541137
• 4,300+ production users of PKBoost library across fraud detection, medical diagnosis, and anomaly detection applications
• Featured on Kaggle with working notebooks demonstrating 86.56% PR-AUC on credit card fraud detection (0.173% fraud rate)

• 100+ GitHub stars across projects (PKBoost, LEMMA, RAG)
• Active maintenance with continuous updates and community engagement
• Mozilla Public License 2.0 and Apache 2.0 licensing for research reuse

• Mentored by Ash Vardanian, industry expert in high-performance vector search and SIMD optimization (USearch used by Anthropic, Cohere, major AI companies)

• Founded PKBoost AI Labs as platform for open ML research
• Regular technical blog posts and documentation for reproducible research

• MMA District Gold Medalist 2022 (5-1 record)
• TRCAC Chess Gold Medalist 2024

Engineering and combat sports share the same DNA:

• Pressure Testing: A system's reliability is only proven when stressed to its limits
• Fundamental Mastery: Deep knowledge of data structures, operating systems, and mathematics over framework hype
• Relentless Improvement: Building elite systems requires daily discipline

I'm open to ML Systems Engineer, Applied AI Researcher, or Performance Engineering roles where technical rigor is the standard.

Email: kharatpushp16@outlook.com
Phone: +91 98696 05981
LinkedIn: pushp-kharat
GitHub: Pushp-Kharat1
Website: pushp-kharat1.github.io

Support My Work: Buy me a coffee ☕

"Build quietly. Measure carefully. Improve relentlessly."