# Krisna Pranav

Senior AI/ML engineer at Turing, based in Singapore. Coding since 6th grade.

Website: https://krishpranav.github.io/
Email: krisna.pranav@gmail.com
GitHub: https://github.com/krishpranav
LinkedIn: https://linkedin.com/in/krishpranav
X: https://x.com/krishpranav5
Résumé: https://krishpranav.github.io/Krisna_Pranav_Resume.pdf

## Summary

I'm Krisna Pranav (krishpranav), a software engineer with 8+ years of production systems and 4+ years in AI/ML. At Turing I build agent infrastructure: agentic workflows and LLM orchestration on Anthropic models, with durable state, so long-running agents resume after a failure instead of starting over. There I work directly with teams at Anthropic, OpenAI, Google and Meta. Before Turing I was forward deployed at Crossing Hurdles, the first engineer at a VC-backed AI startup, worked in Uber's AI labs on work used to train frontier AI models, and moved PayPal services handling 15 to 20 million requests a day from C++ to Go with no downtime.

I'm based in Singapore, where I work, and I travel often to Bengaluru, India. I've also been to San Francisco.

I started out as an ethical hacker, a white hat. The first language I ever learned was Scratch, blocks snapped together on screen. From there it was one project at a time, years before any of it became a job.

## Selected work

- **Turing: I build agents that pick up where they failed.** Agentic workflows on Anthropic models, with durable state, so a long-running agent resumes after a failure instead of starting over. Underneath sit a gRPC routing layer and hot paths in Rust and C++, bridged to Python.
- **Stealth AI notetaker: I was the first engineer, from the first commit.** A VC-backed, privacy-first AI meeting assistant. One bot joins every call, and LangGraph agents turn live transcripts into summaries, action items and decisions.
- **PayPal: Nothing went down.** 15 to 20M requests a day, moved from C++ to Go. Goroutines with bounded channels for backpressure, and a cutover that shifted traffic over in stages, with p99 latency on par with the legacy system.
- **Uber: My work at Uber’s AI labs helped train frontier models.** Alongside it, model-driven services that make real-time decisions on high-traffic surfaces, held to tight latency budgets, and the Go and Python backends that serve them.
- **Crossing Hurdles: I kept first tokens fast, under real load.** Forward deployed inside the teams at Handshake, Otter and Snorkel: a streaming gateway over WebSockets and SSE, and an embedding cache that ended full re-indexes.

## Experience

### AI/ML Engineer, Turing (Now, current)

Agentic AI systems, LLM orchestration, production AI infrastructure.

- Agentic workflows and LLM orchestration on Anthropic models: multi-step reasoning, tool calling, and durable state that lets long-running agents resume after a failure instead of starting over.
- The gRPC/REST layer that routes requests to agents, runs tasks concurrently, and defines the data contracts between services, so new workflows plug in without changes to the core.
- Latency-sensitive infrastructure in Rust (async/Tokio) and C++, bridged to Python through PyO3/FFI for model integration and inference serving.
- Rubric-based evaluation of model outputs, plus inference tuning, observability, and deployment for frontier-model work delivered to Anthropic, OpenAI, Google, and Meta.

Stack: Rust, C++, Python, gRPC, Anthropic.

### AI/ML Engineer, Crossing Hurdles (Mar 2026 to Sep 2026)

Forward deployed on production AI for Handshake, Otter, and Snorkel.

- Worked inside client engineering teams and closed every scoped task on their real-time inference and orchestration paths, in Python, Rust, C++, and TypeScript.
- A streaming inference gateway that fans out concurrent LLM calls and streams tokens back over WebSockets/SSE, keeping time-to-first-token low under concurrent load.
- An in-memory embedding cache and incremental re-indexing, so retrieval stayed fast as traffic grew and content updates no longer triggered a full re-index.

Stack: Python, Rust, C++, TypeScript, SSE.

### Founding Engineer (#1), Stealth AI notetaker (Jan 2025 to Jul 2026)

VC-backed, privacy-first AI meeting assistant, built from zero to one.

- Made the architecture calls and shipped the product from the meeting bot to the user-facing app: Python/FastAPI services on PostgreSQL, Dockerized and deployed on AWS.
- A bot that joins Zoom, Microsoft Teams, Google Meet, and phone (PSTN) calls on a shared SDK layer. Supporting a new platform meant writing one adapter, not another bot.
- Live transcription streams into LangGraph agents that produce structured summaries, action items, and decisions, with sensitive meeting data kept under the user's control.
- RAG over each user's meeting history, with Langfuse tracing on every LLM call so latency, token cost, and bad outputs trace to a specific prompt and step.
- CI/CD, the automated test suite, and code-review conventions; mentored the engineers who joined next and worked directly with early users.

Stack: Python, FastAPI, PostgreSQL, LangGraph, AWS.

### AI/ML Engineer, Uber (Mar 2024 to Aug 2026)

AI labs: work used to train frontier AI models.

- Worked in Uber's AI labs, on work used to train frontier AI models.
- Built and integrated model-driven services that make real-time decisions on high-traffic surfaces, working to tight latency budgets and reliability targets.
- Data and inference pipelines behind ML features, and the Go/Python backends that serve predictions, alongside data science, infrastructure, and product teams.

Stack: Go, Python.

### AI/ML Engineer (Contract), Legal RAG platform (Jun 2021 to Jan 2022)

Retrieval-augmented legal and compliance information system.

- A production RAG system answering procedural, compliance, and legal questions from a curated knowledge base. Every answer cites its source passage, so reviewers can check it.
- The ingestion pipeline: PDF parsing, semantic chunking, metadata tagging (source, jurisdiction, document type), and embedding upsert into a hybrid vector + keyword store with reranking. Scanned pages go through vision-LLM extraction.
- A scheduled monitoring agent that watches designated sources, detects new rulings, and re-indexes the live knowledge base, which ended manual upkeep of the corpus.

Stack: Python, RAG, Hybrid search, Vision-LLM.

### Senior Rust Engineer (Contract), Aether, AI desktop agent (Sep 2020 to May 2021)

On-device retrieval and vector search.

- A privacy-first desktop assistant in Rust: a local semantic engine indexes files, embeds them on-device, and answers context-aware queries with low latency.
- The on-device HNSW vector index, for strong recall on large corpora without a GPU; owned packaging and release of the macOS/Linux binaries.

Stack: Rust, HNSW, Tauri.

### Senior Software Engineer (External Contract), PayPal (Aug 2019 to Jul 2020)

High-throughput backend migration from C++ to Go.

- Migrated critical C++ services to Go on a tier handling 15 to 20 million requests/day, matching the legacy system's p99 latency while simplifying deploys.
- Transaction-processing concurrency on goroutines with bounded channels for backpressure; ran the cutover with zero downtime by shifting traffic over gradually.

Stack: Go, C++.

### Contracts, Earlier roles (2017 to 2019)

Trading infrastructure, full-stack web, mobile and blockchain.

- A low-latency on-chain trading platform (Rust, Next.js/TypeScript).
- Full-stack web (Next.js, React, Node.js, Python).
- Production mobile and blockchain apps (Flutter, React Native, Solidity).

Stack: Rust, TypeScript, Flutter, Solidity.

## Open source

Fixes merged upstream:

- **tauri-apps/tauri #15224** (https://github.com/tauri-apps/tauri/pull/15224): Scoped Retained::as_ptr() bindings replace Retained::into_raw(), so WebContent processes are freed on macOS and iOS. RSS held flat over a 3-minute stress loop.
- **microsoft/playwright #41986** (https://github.com/microsoft/playwright/pull/41986): Opening chrome://extensions from a non-persistent CDP context segfaulted Chromium. A per-browser blocklist turns the crash into an actionable error, with 7,800+ tests passing.

Own projects:

- [vesper](https://github.com/krishpranav/vesper): A username OSINT tool, in Rust. (335 stars)
- [socialosint](https://github.com/krishpranav/socialosint): Finds a target’s emails published on Instagram, LinkedIn and Twitter, and checks them against PwnDB for leaked credentials. In Rust. (101 stars)
- [nexa](https://github.com/krishpranav/nexa): A cross-platform UI framework from one codebase, with reactive signals, streaming SSR and GPU rendering. (64 stars)
- rust-ui/ui: 4 merged PRs, Rust ports of UI components plus fixes.

In the last year: 1,712 contributions on GitHub, across 508 public repositories.

## How I work

I specialise in production AI systems.

Eight years shipping production systems, four of them in AI/ML. I build the infrastructure under agents, the orchestration, the evals and the tracing, not a thin script on top. A three-line LLM demo looks great for thirty seconds. The engineering starts when it meets real traffic: rate limits, slow tools, outputs that change between runs, thousands of concurrent users. I build for that part, and I write code other engineers can build on.

- Build the platform side: Distributed backends, APIs, and cloud-native services under the model, not just prompts on top.
- Trace everything: Every LLM call traced, so a bad output points to a specific prompt and step.
- Design for the next engineer: Adapters and data contracts, so new work plugs in without touching the core.

## Skills

Languages: Python, Go, TypeScript, Rust, C / C++, SQL.

- AI and agents: LangGraph, LangChain, Claude / Anthropic, OpenAI, RAG with hybrid reranking, FAISS / HNSW.
- Distributed systems: Concurrency and backpressure, asyncio / Tokio, WebSockets / SSE, Zero-downtime migrations, Latency profiling.
- Cloud and MLOps: AWS, Azure (AKS / ACI), Docker, Kubernetes, GitHub Actions, Langfuse evals.
- Backend and data: FastAPI, gRPC, REST, PostgreSQL, Redis, Neo4j.

## Off the clock

- 6,250 hours in CS:GO, and 1,250 hours in Fortnite.
- I’m a professional sim racer.
- I trade crypto, meme coins and the stock market.

## Cities

- Singapore: Where I’m based and work.
- Bengaluru: Where I travel often.
- San Francisco: California.

Last updated: 2026-11-02.