Anand Kamble

AI engineer & researcher
Based in San Francisco

Anand Kamble · Founding engineer at Liebre.ai

Anand Kamble. Ideas intoworkingsystems.

I build AI agents, developer tools, and machine-learning systems.

Explore my work View résumé
Engineering × research × craftScroll to explore Selected work, 2022 — 2026

01AI engineering projects

Built to dosomething real.

AI agents, the tools around them,
and the research underneath.

NightShift / Agent orchestration02
01Worker

The work begins.

An autonomous coding agent takes on a ClickUp ticket in its own PTY-backed session.

02QA

A second look.

The work moves through a dedicated QA stage, coordinated by lifecycle hooks and custom prompts.

03Review

From ticket to pull request.

A PR-review stage completes the workflow, keeping coding, QA, and review as distinct responsibilities.

A process, not a single prompt.Explore the stages

Agent infrastructure · Liebre.ai

NightShift.

Good work doesn’t
happen in one step.

Built a harness from scratch to run autonomous Claude Code agents on ClickUp tickets through worker, QA, and PR-review stages. Custom prompts, lifecycle hooks, and PTY-backed terminals coordinate agent sessions.

  • TypeScript
  • React
  • xterm.js
  • Claude Code

Further explorations

03

An MCP server that lets AI assistants interact with Instagram through the Model Context Protocol.

  • Model Context Protocol
  • AI assistants
04

Applied language models

Automatic Grading System

Developed an assignment-grading platform using fine-tuned language models and retrieval-augmented generation, with prompt engineering to improve grading accuracy and reduce hallucinations.

  • Fine-tuning
  • RAG
  • Prompt engineering
05

Domain-specific machine learning

Aircraft Fault ATA Prediction

Fine-tuned Phi-3-mini on aircraft maintenance logs to predict ATA codes using LoRA adapters and 8-bit mixed precision. Built a configurable pipeline for processing new logs.

  • Phi-3-mini
  • LoRA
  • 8-bit mixed precision

02AI research & publications

A little deeper.A little further.

Publications & preprints on speech, language models, and applied machine learning.

Google Scholar
  1. 01

    arXiv:2607.05623 / 2026 · Preprint

    NAVER LABS System Re-implementation for the IWSLT 2026 Instruction-Following Task

    Anand Kamble, Aniket Tathe

    A speech instruction-following system combining SeamlessM4T-v2-large with Qwen3-4B-Instruct. The work adapts a three-stage training pipeline and builds synthetic examples for ten speech tasks.

  2. 02

    arXiv:2403.00212 / 2024 · Preprint

    Transcription and translation of videos using fine-tuned XLSR Wav2Vec2 on custom dataset and mBART

    Aniket Tathe, Anand Kamble, Suyash Kumbharkar, Atharva Bhandare, Anirban C. Mitra

    Hindi video transcription and translation using a voice-adapted speech recognition model and mBART. Retrieval-Based Voice Conversion helps build training data from a short audio sample, with translations aligned to the video timeline.

  3. 03

    arXiv:2401.06183 / 2024 · Preprint

    End to end Hindi to English speech conversion using Bark, mBART and a finetuned XLSR Wav2Vec2

    Aniket Tathe, Anand Kamble, Suyash Kumbharkar, Atharva Bhandare, Anirban C. Mitra

    A Hindi-to-English speech pipeline connecting automatic speech recognition, neural machine translation, and speech synthesis to turn spoken Hindi into English audio.

  4. 04

    arXiv:2311.14836 / 2023 · Preprint

    Custom Data Augmentation for low resource ASR using Bark and Retrieval-Based Voice Conversion

    Anand Kamble, Aniket Tathe, Suyash Kumbharkar, Atharva Bhandare, Anirban C. Mitra

    Two approaches to preparing voice-specific training data for low-resource speech recognition: audio generation with Bark and voice conversion with RVC, applied to customized Common Voice datasets.

  5. 05

    Materials Today: Proceedings (Elsevier) / 2022

    Object following robot based on AI/ML

    Anand Kamble, Anirban C. Mitra, Aniket Tathe, Suyash Kumbharkar, Atharva Bhandare

03About Anand Kamble

Curiosity,put to work.

At Liebre.ai, I work on Luca, an AI accounting agent, and NightShift, a harness for autonomous coding agents. My background spans language-model research, high-performance computing, and frontend engineering.

Experience

May 2025 – PresentLiebre.aiFounding Engineer
  • Built LLM abstraction layers with versioned prompt management and OpenTelemetry tracing, and deployed LLM/ML microservices to improve time to first token.
  • Designed transaction clustering with text embeddings, UMAP, and categorical features, achieving approximately 95% cluster purity.
  • Implemented real-time agent streaming with SSE and WebSocket events, a Redis-backed event bus, and JSON Patch state synchronization.
Jan 2024 – Oct 2024Florida State UniversityGraduate Research Assistant
  • Developed retrieval-augmented generation pipelines, secure prediction APIs, and multi-agent workflows using hosted and self-hosted language models.
  • Designed and deployed model-parallel inference on a SLURM-managed HPC cluster, achieving an 8× speedup.
  • Evaluated model performance on synthetic datasets and iterated on prompt engineering.
Apr 2022 – Nov 2022IncerroAIFrontend Engineer
  • Built responsive web applications with real-time data binding.
  • Reduced React re-renders by 40% through memoization and efficient state management.

A working toolkit

Agents & orchestration
Agent architecture · Claude Code / pi harnesses · MCP · LangGraph · LangChain · LlamaIndex
Languages
Python · TypeScript · Java · C/C++ · Bash · SQL
Applications & machine learning
FastAPI · React · Express · PyTorch · scikit-learn · llama.cpp · LiteLLM
Cloud & infrastructure
GCP: Cloud Run & Cloud SQL · AWS · Docker · Kubernetes · PostgreSQL · Redis · SLURM
Testing & delivery
PyTest · Vitest · Playwright · Testcontainers · GitHub Actions
Tools & observability
Git · Linux/Unix · OpenTelemetry · JIRA

04Make contact

What’s thenext idea?

Write to me at

hey@anandkamble.com
LinkedIn ↗GitHub ↗