Medhaja H L · Bengaluru, India

I build AI products from idea to deployment.

I’m a GenAI Technical Lead and AI product builder, combining product thinking, full-stack engineering, and production GenAI systems. I’m also the creator of Isocline, an open-source platform for building agent workflows.

Open to Staff / Lead GenAI opportunities and consulting engagements.

Explore my workGet in touch

WorkFeatured product

Isocline

An open-source control plane and durable execution runtime for agent workflows, with a visual builder on top.

  • Open source · Apache-2.0
  • Creator & developer
Isocline desktop app: a research workflow where specialist agents run in parallel, merge their outputs, and wait at a human-approval step before generating a report.
A demo run paused at the human-approval step.
ready-to-run templates
53
src: docs/verification/release-3.3.0.txt
tests passed
233
src: docs/verification/release-3.3.0.txt
durability scenarios
10/10
src: docs/verification/release-3.3.0.txt
model providers
6+
OpenAI · Anthropic · Gemini · OpenRouter · Ollama · OpenAI-compatible
The problem
Building an agent is becoming easy; operating agent workflows reliably is not. Model calls fail and rate-limit, workers crash mid-run, approvals take days, side effects need undoing, and generated code has to run somewhere safe.
Who it’s for
Developers and teams who want to compose agent workflows visually and run them on their own infrastructure with multiple model providers or any OpenAI-compatible endpoint. A Windows desktop build and 53 ready-to-run templates also make it usable by individuals.
My role
I created Isocline and develop it: architecture, runtime, web builder, SDK and CLI, desktop packaging, and release verification.

Demonstrated workflow

A research workflow runs specialist agents in parallel, merges their outputs, and pauses for human approval before generating a report.

This is demonstrated functionality, not a customer deployment.

Architecture, layer by layer Every tier as its own plate. PostgreSQL holds the authoritative state, so Redis can be lost without losing a run. Concept drawing. Sizes are illustrative.Read the architectureOpen the interactive drawing
Exploded concept drawing of Isocline's architecture. Top to bottom: the Web UI and SDK/CLI clients; the FastAPI API; PostgreSQL with pgvector, highlighted as the authoritative state, beside Redis; Celery workers and Beat; and model providers, tools and MCP servers, and the Python sandbox.
Isocline — demo dashboard
Isocline dashboard: summary tiles for runs, success rate, tokens, spend and average duration; a table of recent completed workflow runs; and a list of models in use across OpenAI, Google and Anthropic.

What it does today

  • Visual builder with typed contracts between nodes, checked before a run starts

  • Durable execution on PostgreSQL: parallel DAGs, crash recovery, checkpoints, rerun from any node

  • Human-in-the-loop waits (approval, timer, callback, event) that hold no worker

  • Saga-style compensation that undoes side effects in reverse order

  • Sandboxed Python in disposable, network-less containers

  • Multiple model providers: OpenAI, Anthropic, Gemini, OpenRouter, Ollama and OpenAI-compatible servers

  • MCP tools, RAG on pgvector, webhook and schedule triggers, REST API, Python SDK and CLI

  • Case study Read the case study
Read the architectureSource on GitHubDocumentationWindows installerTutorial video

Selected work

Voice AI · Agent orchestration

L1 Interviewer Bot

A voice-first conversational agent that screens technical candidates, using RAG to probe specific technical domains, with automated technical and communication scoring and a low-latency voice-response loop.

Demo video: L1 Interviewer Bot

Earlier ML work

Smaller projects with public code

Personal learning projects from before Isocline. They are experiments, not production systems.

2024 · LLM evaluation

Evaluating a RAG question-answering pipeline with MLflow

Problem
Measure the answer quality of a retrieval-based QA pipeline instead of judging outputs by eye.
What I did
Built a LangChain RetrievalQA pipeline over a PDF with a FAISS index, then scored it with mlflow.evaluate using LLM-judged faithfulness and relevance metrics plus latency. Packaged with Docker.
Stack
Python · LangChain · FAISS · OpenAI · MLflow · Docker
Learning
Wiring LLM-as-judge metrics into experiment tracking. The evaluation set is four questions, so this is a pattern exercise, not a benchmark.
Source: Evaluating a RAG question-answering pipeline with MLflow

2024 · Tabular ML

Heart failure prediction with XGBoost and MLflow

Problem
Predict a death event from 13 clinical features in a public Kaggle dataset of heart-failure records.
What I did
Trained an XGBoost classifier, ran a 50-trial Hyperopt search with every trial logged as a nested MLflow run, and evaluated with accuracy, ROC AUC, a confusion matrix and feature-importance plots.
Stack
Python · XGBoost · Hyperopt · scikit-learn · MLflow
Learning
Using MLflow to make hyperparameter search reproducible and comparable across trials.
Source: Heart failure prediction with XGBoost and MLflow

2023 · Computer vision

In-browser image classifier with transfer learning

Problem
Classify two product types from a webcam image entirely on the user’s device, without sending images to a server.
What I did
Fine-tuned a VGG16 model, converted its weights for TensorFlow.js, and built a static page that classifies webcam captures in the browser.
Stack
TensorFlow / Keras · VGG16 · TensorFlow.js · static hosting
Result
A working demo that runs on-device. The demo asks for webcam access.
Live demo: In-browser image classifier with transfer learningSource: In-browser image classifier with transfer learning

Experience

Where I’ve worked

5+years of industry experience

  1. – Present

    GenAI Technical LeadCurrent

    Biz-Metric India Pvt. Ltd.

    • Led development of an enterprise agentic accelerator that migrates Azure Data Factory (ADF) pipelines to PySpark, combining deterministic translation, LLM-based semantic validation and automated repair workflows, reducing manual migration effort by 90%.
    • Lead delivery of enterprise document-processing and knowledge-retrieval applications, owning the technical architecture and guiding engineers through implementation, evaluation and deployment.
    • Build document-processing agents and RAG pipelines, orchestrating multiple models with LangGraph and LiteLLM, including routing, failure handling and output-quality checks.

  2. –

    Senior Software Engineer

    Infinite Computer Solutions

    • Deployed Mistral-7B with vLLM on Google Kubernetes Engine (GKE) for GPU-backed inference serving, and explored speculative decoding to improve serving performance.
    • Built FastAPI services and evaluation components for enterprise AI inference and evaluation systems, integrating LLM and OpenCV components into evaluation workflows.
    • Built a multi-agent proof of concept with AutoGen.

  3. –

    Software Engineer, AI/ML

    SourceFuse

    • Built enterprise knowledge-retrieval (RAG) applications with FAISS and FastAPI, connecting document retrieval to LLM generation.
    • Integrated multiple model backends — OpenAI, Mistral, Llama 2 and Amazon Bedrock — into these applications.

  4. –

    Analyst

    EXL

    • Developed NLP question-answering solutions and PySpark data-processing workflows, and worked on Docker and Kubernetes deployments.

  5. Earlier

    • –

      Machine Learning Engineer (Freelance) PyImageSearch

      Converted Python projects into cloud-based notebooks and wrote blog posts on the process.

    • –

      Machine Learning Engineer Intern Harman International

      Built anomaly detection on real-time 5G telematics data using autoencoders on Amazon SageMaker.

Capabilities

What I work on

GenAI systems

RAG, agent workflows, tool use, structured extraction, LLM evaluation.

Evidence:

Application engineering

Python, FastAPI, APIs, React, end-to-end application development.

Evidence:

Inference and deployment

vLLM, Docker, Kubernetes, cloud deployment.

Evidence:

Reliability

Tracing, failure handling, evaluation pipelines, latency and cost analysis.

Evidence:

Product and leadership

Problem definition, prototyping, technical design, delivery, mentoring.

Evidence:

About

Engineer, technical lead, product builder.

I’m a GenAI Technical Lead based in Bengaluru, with a background in Computer Science and Engineering from MIT Manipal.

My work combines hands-on engineering, technical leadership, and product development. I enjoy taking an ambiguous business problem, building a useful prototype, and developing it into a dependable application.

Alongside my professional work, I build Isocline. I’m interested in opportunities where I can own AI products, work closely with users, and lead delivery from concept to deployment.

Background

Education & Certifications

Education

B.Tech in Computer Science and EngineeringManipal Institute of Technology, Manipal2015 – 2019

Certifications

Work with me

Let’s build something useful.

For hiring

Staff / Lead GenAI roles

Interested in Staff / Lead GenAI and hands-on AI product engineering opportunities.

For consulting

GenAI applications

I help teams prototype and build AI workflow automation, knowledge assistants, and GenAI applications.

Emailmedhajahlbhat@gmail.com