Anurag Pal | Chartered Accountant & Software Engineer

About Me

I am a Chartered Accountant with an enduring passion for serious systems programming, AI engineering, and full-stack software development.

My engineering journey began on the audit floor. Frustrated by traditional, high-risk sampling methods, I built custom automated tools—including **3-way inventory & purchase reconcilers**, **WMS dispatch vs. stock ledger checkers**, **creditor aging pipelines**, and **TDS tax compliance validators**—enabling 100% data verification across millions in transactions.

Driven by what code and mathematics can achieve, I expanded into deep learning and natural language processing, constructing **GPT-style Transformers** and **Byte-Pair Encoding (BPE) tokenizers from first principles**, as well as **private local LLM extraction pipelines with JSON schema enforcement** for earnings calls.

Today, I specialize in **systems programming in Rust**—engineering high-performance autonomous agent runtimes with dynamic planning and semantic context compaction, native desktop apps with Slint, and cross-platform web & mobile systems.

I bring a rare convergence of **rigorous financial & domain understanding** and **low-level systems & AI engineering**, crafting robust, deterministic software for mission-critical operations.

What I Deliver

Financial & Audit Automation

Bespoke reconciliation engines, automated 3-way invoice/GRN/ledger matching, Tally ERP integrations, and tax compliance automation (TDS) that eliminate 100s of manual spreadsheet hours.

Private Local AI & Agent Runtimes

Self-hosted, on-premise local LLMs (Ollama) and autonomous agent harnesses with strict JSON schema enforcement to parse earnings calls and financial filings with zero cloud data leakage.

Web & Cross-Platform Mobile

Full-stack server-side rendered (Next.js) web applications and cross-platform mobile apps (React Native Expo) powered by unified TypeScript APIs and robust PostgreSQL databases.

High-Performance Native Systems

Ultra-fast native desktop applications (Rust + Slint UI), high-throughput data processing pipelines, and deterministic CLI developer tooling with minimal memory footprint.

Case Studies & Production Solutions

Production Engine

3-Way Inventory & GRN Reconciler

Automated matching pipeline cross-verifying ERP purchase registers, warehouse Goods Receipt Notes (GRN), and physical stock ledgers. Detects rate discrepancies, unrecorded liabilities, and inventory shrinkage automatically.

Node.js Python WMS Integration Internal Controls Variance Analysis

Privacy-Safe AI

Private Local LLM Financial Intelligence

On-premise financial extraction pipeline running local SLMs (Llama 3 / Qwen via Ollama) with strict JSON schema constraints to extract metrics, guidance, and risks from earnings call transcript PDFs without cloud API costs or data leakage.

Python Ollama Local LLMs JSON Schema Financial Analytics

Compliance Pipeline

Creditor Audit & TDS Compliance Suite

Automated batch verification tool for vendor ledger aging, statement confirmations, and Indian Tax Deducted at Source (TDS) rate validation, flagging statutory non-compliance and unbilled liabilities before annual closing.

Python Tax Auditing TDS Compliance Tally ERP / TDL Automation

Dual-Stack App

Unified Web & Mobile Goal Ecosystem

Production-grade productivity platform featuring a Next.js SSR web application and a synchronized React Native Expo mobile app, sharing a unified TypeScript backend and authentication architecture.

Next.js React Native Expo TypeScript REST APIs PostgreSQL

Featured Open Source Projects

Autonomous Rust Agent Runtime

High-performance local AI agent runtime engineered in Rust with Tokio. Features 3-phase orchestration (Plan, Execute, Synthesize), proactive token compaction (reducing 14+ messages to 5), reasoning fallback for local SLMs (Qwen/DeepSeek), and sandboxed execution.

Rust Tokio Async Ollama / Local LLMs Agentic AI

Nano-GPT Transformer from Scratch

Complete ground-up implementation of a decoder-only GPT Transformer architecture in PyTorch. Implements multi-head causal self-attention, residual connections, layer normalization, position embeddings, and autoregressive text generation.

Python PyTorch Transformers Deep Learning NLP

Byte-Pair Encoding Tokenizer

Custom BPE tokenizer built from scratch in Python following first principles. Implements byte-level vocabulary training, pair frequency extraction, merge rule persistence, and optimized encoding & decoding pipelines for LLMs.

Python Tokenization BPE LLM Foundations

Audit Sampling Risk Calculator

A specialized financial auditing tool built to calculate statistically sound sample sizes for internal control testing and substantive tests of details based on population size, risk levels, tolerable error, and expected deviations.

JavaScript Audit Analytics Internal Controls FinTech

Native Desktop GUI Application

High-performance cross-platform native desktop productivity app engineered with Rust and the Slint UI declarative GUI framework. Features zero-overhead memory footprint, fluid animations, and local persistence.

Rust Slint UI Desktop GUI Systems

Next.js SSR Content Platform

Full-stack server-side rendered (SSR) web application built with Next.js, React, and TypeScript. Optimized for Core Web Vitals, dynamic routing, SEO performance, responsive layouts, and lightning-fast edge delivery.

Next.js TypeScript React Tailwind CSS Vercel