# Krrish Choudhary

> Krrish Choudhary is a software and AI engineer who builds AI assistants and the backends that run them.

**I build AI assistants and the backends that run them.**

I'm Krrish Choudhary, a software and AI engineer finishing a B.Tech in Computer Science at LNMIIT Jaipur in 2027. At Incredible I shipped an AI agent platform that connects **300+ apps** and handles **about 10 million events a month**. At GoQuant I took **120 microseconds** off a trading data path that already ran in under a millisecond. I now run Verace, where I'm building a new LLM architecture and open-source memory tools for AI coding agents.

- [Resume (PDF)](https://www.krrishchoudhary.in/Krrish_Choudhary_Resume.pdf)
- Email: [krrishchoudhary109@gmail.com](mailto:krrishchoudhary109@gmail.com)

## Work

Each line states the result first; the indented line under it says how it was built.

### Verace

**Founder & Engineer**, April 2026 to now, Remote, India, [verace.in](https://verace.in)

- Founded Verace to build a new LLM architecture and open-source memory tools for AI coding assistants. Working alone, I finished **1,579 tasks across 14 sprints** and launched the company's web platform.
  - How: The web platform at verace.in runs on TypeScript and Next.js 15. The two main products, Verace V1 and Agentic Memory, are described under Projects.
- Designed the cloud backend that serves the model.
  - How: gRPC microservices behind a GraphQL gateway on Kubernetes (AWS EKS and GCP GKE, provisioned with Terraform), Kafka for event streaming, PostgreSQL with pgvector as the model's memory store, and vLLM and Hugging Face serving monitored with Prometheus, Grafana and OpenTelemetry.

### Incredible

**Software Developer**, August 2025 to March 2026, Remote, [incredible.one](https://www.incredible.one)

- Shipped Incredible Chat, an AI assistant that works across **100+ connected apps** such as Slack, LinkedIn and Google Workspace, and made the information it retrieves **about 30% more relevant** in production.
  - How: Retrieval-augmented generation (embedding-based semantic search) plus custom LLM fine-tuning. The gain was confirmed with offline evals and online A/B tests.
- Built the trigger system that lets Incredible's agents start work on their own, on a schedule or when something happens in a connected app, across **300+ services** and **about 10 million events a month**.
  - How: LangChain, MCP (Model Context Protocol) and OpenAI and Anthropic function calling, with webhooks, timezone-aware cron scheduling, OAuth2 account management and at-least-once delivery.
- Raised the share of tasks the agent finishes by **about 45%** over the single-agent baseline.
  - How: A dual-agent LLM architecture that turns a plain-language request into an executable Python workflow across 100+ integrated services and custom agent tools.
- Built the core of the Incredible Voice Assistant, so people can call the assistant by voice from any application in **8 locales**.
  - How: Written in Rust: voice-command activation, on-device natural language processing, and real-time capture of what is on the user's screen so the assistant has context.
- Cut the time it takes to deploy by **about 70%** across **12+ services**.
  - How: Rebuilt the CI/CD pipelines on GitHub Actions with layered build caching, parallel test stages and Helm-managed releases on Kubernetes.
- Delivered the Incredible API platform so outside developers can build on the company's AI.
  - How: Endpoints for agentic reasoning, deep research, web search, image and video generation and OCR, plus a developer portal with API keys, usage analytics and SDK code generation.

### Kanchan

**Software Developer**, May to July 2025, Bhilwara, India

- Cut the median page-load time from **3.2 seconds to 1.4 seconds** on mobile and desktop.
  - How: A CDN-cached asset pipeline with image optimization, fingerprinted bundles and edge caching, which improved Largest Contentful Paint by about 56%.
- Lifted user engagement **about 40%** by leading a full website revamp across three partner teams.
  - How: Coordinated content, design and operations partners and shipped a redesigned React front end with a modular component architecture.

### Vaidsys Technologies

**Software Developer Intern**, May to June 2025, Remote, [vaidsys.com](https://vaidsys.com)

- Shipped a task-management app that stayed fast with **100+ people** using it at once, answering in **about 180 ms** at the 95th percentile.
  - How: An Angular and TypeScript front end with RxJS state management, on a Node.js and Express REST backend with MongoDB.
- Made sure each person sees only the data their role allows, and gave the partner team a way to test every release before it ships.
  - How: JWT authentication with role-based access control (RBAC) on every endpoint, test-first CI checks on GitHub Actions, and a Playwright end-to-end suite covering the critical task flows.

### GoQuant Technologies

**Back-End Engineer Intern**, February to April 2025, Miami, FL (remote), [goquant.io](https://goquant.io)

- Raised throughput **15%** and cut worst-case latency by **about 120 microseconds** on a trading engine's market-data path that already ran in under a millisecond.
  - How: Profiled the hot path of the order book and moved it to allocation-free data structures in low-latency C++, on live Binance and Deribit WebSocket feeds. The latency figure is p99 tick-to-trade.
- Made two exchanges look like one feed to the trading engine.
  - How: Feed adapters that normalize L2 order-book and trade-tick formats across Binance and Deribit, with WebSocket reconnection, heartbeat handling and gap detection.

## Projects

### Agentic Memory

An open-source memory server that lets AI coding assistants such as Claude Code, Cursor and Codex CLI remember what they learned between sessions.

How it was built: A Model Context Protocol (MCP) server with six memory layers behind one ranked query, storage that is safe against prompt injection, and bitemporal belief revision that keeps a full audit history and never overwrites a past belief. It runs on multi-tenant PostgreSQL and pgvector with Row-Level Security, ships as a Claude Code marketplace plugin with 9 autonomous hooks, and is covered by **98 automated tests**.

- [Code on GitHub](https://github.com/Verace-Pvt-Ltd/agentic-memory)

### StreamRAG v2

A live map of a codebase for AI code editors. With it, the editor's AI used **47% fewer tokens** and needed **19% fewer back-and-forth turns** to do the same work.

How it was built: A real-time incremental code graph exposed as an MCP server, with native parser daemons for 8 languages (Python, TypeScript, JavaScript, TSX, Rust, C++, C and Java). A 10-stage Delta GraphBridge pipeline updates the graph in under 0.05 ms per change across 20,000+ lines.

- [Read the paper](https://journal.ijtrp.com/index.php/ijtrp/article/view/15)
- [Open the live demo](https://streamrag-frontend.vercel.app/)

### Hybrid GraphRAG for legal citations

Search that takes a legal question in English and finds the right Swiss citations across **2.47 million** court passages written mostly in German. It can only return citations that exist, so it cannot invent a source. On the validation set it scored **111% higher** than plain keyword search (Macro F1 0.691 against 0.327).

How it was built: Hybrid retrieval that fuses BM25, BGE-M3 embeddings on FAISS, and graph signals from a citation graph (Personalized PageRank and Leiden communities) with Reciprocal Rank Fusion, followed by a BGE-reranker-v2-m3 cross-encoder and an agentic Qwen2.5-7B verifier. It runs on two T4 GPUs. Co-authored with T. Kandoi and S. Patel.

- [Read the paper](https://doi.org/10.18535/ijecs/v15i03.5461)
- [Code on GitHub](https://github.com/Krrish2004/legal-graphrag)

### DriftCall

A training ground that teaches a voice-first booking assistant (flights, cabs, food, hotels, payments in Indian languages) to notice when the APIs it relies on change underneath it. After training, it caught **65 percentage points more** of those changes.

How it was built: An OpenEnv-compliant reinforcement learning gym. A drift engine mutates 5 versioned mock APIs in 20 patterns (schema, policy, terms, pricing, auth) to generate 200,000+ episodes, scored by 5 deterministic rewards with Brier calibration. Gemma 4 was fine-tuned with LoRA (r=16) and GRPO through Unsloth and TRL for 500 steps on one V100. Speech goes in through faster-whisper and out through Kokoro, behind a FastAPI loop.

- [Code on GitHub](https://github.com/Krrish2004/DriftCall)

### Verace V1

A text-and-vision LLM architecture designed from scratch so that memory per token stays constant and each token uses only as many layers as it needs, from **2 to 128**.

How it was built: It replaces softmax attention, the KV-cache and routed mixture-of-experts with continuous manifold-based mechanisms, and adds a latent energy critic that picks the lowest-energy branch during tree-search generation. This is a reference implementation in PyTorch with custom Triton GPU kernels and a 7-module unit and integration test suite; it has no pretrained weights yet.

- [Code on GitHub](https://github.com/Verace-Pvt-Ltd/Verace-V1)

### NeurIPS 2025 Open Polymer Prediction

Predicts five physical properties of a plastic, such as its density and how well it conducts heat, from its chemical formula alone, so new materials can be screened on a computer before anyone makes them. It learns from **7,973 molecules** even though four of the five properties are missing for **more than 90%** of them.

How it was built: An ensemble of two models: a 6-layer Transformer (8 heads, 256 dimensions) that reads the formula as a SMILES string, and a 3-layer graph convolutional network that reads the molecule as a graph of atoms. A custom SMILES-to-graph pipeline with no RDKit dependency supplies a 43-token tokenizer, 19 features per atom and 29 molecular descriptors, with KNN imputation. Multi-task training uses a masked loss, so each molecule teaches only the properties it has, tuned with AdamW, ReduceLROnPlateau and Optuna search in PyTorch on CUDA. The five targets are glass transition temperature, fractional free volume, thermal conductivity, density and radius of gyration; weighted MAE was 8.418 in 5-fold cross-validation.

- [Code on GitHub](https://github.com/Krrish2004/NeurIPS-Open-Polymer-Prediction-2025)

### Earlier projects

- [Ubuntu Time Machine](https://github.com/Krrish2004/ubuntu-time-machine): Incremental, encrypted backups for Ubuntu with a timeline to browse old snapshots. C++ core, Electron and React interface.
- [Quantum key distribution for payments](https://github.com/Krrish2004/qkd-razorpay-demo): A demo that protects a Razorpay payment flow with keys from a simulated BB84 quantum protocol, plus a fraud check. Qiskit, PyTorch, Flask.

## Papers and writing

- [Hybrid GraphRAG for Cross-Lingual Legal Citation Retrieval: A Multi-Signal Fusion Approach for Swiss Legal Information Systems](https://doi.org/10.18535/ijecs/v15i03.5461). IJECS, volume 15, issue 4, April 2026. With T. Kandoi and S. Patel.
- [StreamRAG v2: Native Parser Daemons and Three-Tier Extraction for Real-Time Code Graphs](https://journal.ijtrp.com/index.php/ijtrp/article/view/15). IJTRP, volume 2, issue 2, February 2026.
- [The Artificial Brain: A Neuroscience-Inspired Architecture for Multimodal AI Systems](https://journal.ijtrp.com/index.php/ijtrp/article/view/13). IJTRP, volume 2, issue 2, February 2026. With T. Kandoi. A design proposal: small specialized models handle inputs in parallel, a larger reasoning model wakes only when needed.
- [The Founder's Atlas: The First Principles Concepts That Decide Company Outcomes](https://www.amazon.in/dp/B0H57GQ4CG). A book for founders, June 2026. A first-principles guide to the 179 concepts that decide how a company turns out, from market sizing and moats to unit economics and growth loops. 682 pages, available on Amazon.
- [Order Flow Trading (XAUUSD), first edition](https://www.amazon.in/Order-Flow-Trading-XAUUUSD-Institutional/dp/B0GHK72C38). A book on reading the order book in the gold market, 2026. Available on Amazon.
- [All papers on Google Scholar](https://scholar.google.com/citations?user=3umhN38AAAAJ&hl=en)

## About

### Education

**LNMIIT, Jaipur.** B.Tech in Computer Science Engineering, August 2023 to May 2027 (expected).

Coursework includes data structures and algorithms, operating systems, computer networks, databases, deep learning and large language models. My capstone project detects deepfakes with graph neural networks.

### Recognition

- **Codeforces Expert**, with a peak rating above 1600. [Profile](https://codeforces.com/profile/krrishchoudhary109)
- **Ranked first in my college** at AlgoUniversity Graphs Camp: 30+ problems in 3 days using 17+ advanced graph algorithms. [Post](https://www.linkedin.com/posts/krrish-choudhary-91b1a8281_achievement-graphscamp-dsa-activity-7295783171260850177-VyLc)
- **Top 50 of 40,000+ applicants**, named to the AlgoUniversity Hall of Fame.
- **LUSIP research intern** at LNMIIT, 2024: trained models that detect AI-generated singing voices with about 85% validation accuracy.

### Tools I use

- **Languages**: TypeScript, JavaScript, Python, Rust, C++, Go, Java, SQL, Bash
- **AI and LLMs**: AI agents, function calling, RAG, fine-tuning (LoRA, GRPO), LangChain, MCP, PyTorch, Hugging Face Transformers, vLLM, OpenAI and Anthropic APIs, speech recognition and synthesis
- **Backend and web**: Next.js, React, Angular, Node.js, Express, FastAPI, REST, gRPC, GraphQL, WebSockets, Kafka, OAuth2, JWT
- **Cloud and data**: AWS, GCP, Kubernetes, Docker, Terraform, PostgreSQL, pgvector, MongoDB, FAISS, GitHub Actions, Linux

### Outside engineering

I trade forex and have written two books: one on order-flow trading in the gold market, which is where my interest in low-latency systems started, and The Founder's Atlas, a first-principles guide for people starting companies.

## Contact

The fastest way to reach me is email. I'm based in Jaipur, India, and I work remotely with teams in other time zones.

- Email: [krrishchoudhary109@gmail.com](mailto:krrishchoudhary109@gmail.com)
- [LinkedIn](https://www.linkedin.com/in/krrish-choudhary-91b1a8281/)
- [GitHub](https://github.com/Krrish2004)
- [Google Scholar](https://scholar.google.com/citations?user=3umhN38AAAAJ&hl=en)
- [Codeforces](https://codeforces.com/profile/krrishchoudhary109)
- [Kaggle](https://www.kaggle.com/krrishchoudhary109)
- [Resume (PDF)](https://www.krrishchoudhary.in/Krrish_Choudhary_Resume.pdf)

---

This Markdown is generated from the portfolio page at [/](https://www.krrishchoudhary.in/). A short index is at [/llms.txt](https://www.krrishchoudhary.in/llms.txt).
