Open to ML / LLM engineering roles

Hi, I'm Karthik

Machine Learning Engineer. I ship production ranking and personalization systems, and build self-hosted LLM platforms end to end.

Bangalore, India

About

I have 5+ years of production ML experience. At Wiser Solutions I built ranking, personalization and product-matching systems, and the A/B testing platform the whole company used to test models in production. Before that I built genomics pipelines at Rutgers and edge computer vision at Neva Innovation Labs.

Today I also run S.I. Chemicals, my family's chemical distribution business. It gives me something unusual for an ML engineer: a real business with real data and real problems (procurement, pricing, regulatory compliance) to build ML systems against. Most of my current projects come from there, and they run on my own Kubernetes cluster.

I'm looking for ML Engineer, Applied Scientist and NLP / LLM Engineer roles at product companies and startups, India-based or remote.

Work Experience

SIC
S.I. Chemicals
Machine Learning Engineer · Owner / General Manager
Aug 2024 – Present
  • Run a chemical distribution business (methanol, IPA, acetone, thinners) serving customers across multiple industries; lead a team of 10 across operations, inventory, pricing and B2B sales.
  • Build ML systems for the business alongside operations: an agentic SDS compliance assistant and a procurement buy-now / wait signal on chemical market data.
  • Run my own infrastructure: a k3s cluster on a dedicated server, deployed with Flux GitOps, which hosts my ML services.
W
Wiser Solutions
Software Engineer (Machine Learning) · San Mateo, CA
Nov 2021 – Feb 2024
  • Built ranking and personalization models for Mobee, Wiser's crowdsourced shopping app: +15% NDCG@10 offline and +10% CTR in production A/B tests.
  • Built the company's A/B testing and experiment-assignment platform, adopted beyond the ML team, so several model variants could run in production without interrupting live traffic; added an experiment dashboard and Airflow metric pipelines over Segment event data.
  • Built the candidate-generation stage of product matching (entity resolution) over a catalog of several million products, using transformer embeddings and ANN retrieval to produce small, high-recall candidate sets.
  • Developed NER-based attribute extraction (brand / model / variant from raw product titles) to sharpen blocking keys and separate near-identical variants.
  • Drove model productionization for the ML team: FastAPI serving, Kubernetes / Helm manifests, AWS resources via Terraform; helped bring in MLflow and DVC.
  • Built a dynamic pricing model for shopper reward points, part of a pricing initiative that drove +12% revenue and +8% conversion.
R
Rutgers University
Graduate Research Assistant, cancer genomics lab
Jan 2021 – Jun 2021
  • Solo-built four genomics pipeline packages (variant calling, ChIP-seq, RNA-seq, single-cell) that the lab used as core infrastructure.
  • Each is a config-driven CLI that pulls controlled-access NIH data (dbGaP, GDC) and runs multi-stage workflows as independently-resourced SLURM jobs on the Amarel HPC cluster.
N
Neva Innovation Labs
Software Engineer, Machine Learning · Bangalore
Feb 2018 – Jul 2019
  • Designed and deployed driver drowsiness detection and glance classification with an improved face detector, running at 15 FPS on edge devices.
  • Led research on running models on low-power hardware: Intel Movidius, Snapdragon SoC, Nvidia Jetson.
  • Built an ensemble (neural nets, SVM, random forest) for driver behaviour detection from vehicle telemetry, and an air-quality forecaster from a home robot's sensors.

Education

Rutgers, The State University of New Jersey
MS, Computer Science · GPA 3.8
2019 – 2021
Sir M Visvesvaraya Institute of Technology
B.Eng, Computer Science · First Class with Distinction
2013 – 2017
Certifications
RAG and Agentic AI (IBM) · ML Engineering for Production / MLOps (DeepLearning.AI)

Skills

ML & LLM

PyTorch TensorFlow scikit-learn Transformers vLLM Constrained decoding Anthropic SDK MCP RAG Ranking & recsys Entity resolution MLflow

Languages & data

Python Go SQL pandas NumPy PySpark Airflow PostgreSQL Redis MinIO

Infrastructure

FastAPI NATS JetStream Docker Kubernetes Helm Terraform AWS Modal Flux GitOps Traefik GitHub Actions Playwright
Projects

Things I'm building

Mostly ML and LLM systems that solve real problems in my chemical business, all running on infrastructure I operate myself.

FastAPI→ NATS→ Go workers→ Qwen3-4B · vLLM In production

ScrapeFlow

Distributed scraping & LLM extraction

A scraping platform that has grown through three production phases: FastAPI control plane, Go workers over NATS JetStream, hardened Playwright against bot detection. Its extraction service runs Qwen3-4B with vLLM and grammar-constrained decoding, so every result is valid structured data, at about $1/day. The whole platform is exposed to LLMs as MCP tools.

Go NATS JetStream vLLM XGrammar Modal MCP k3s
question→ agents→ SDS · PubChem · REACH→ cited answer Built

SDSage

Chemical compliance agent

A multi-agent system with persistent memory that answers safety-data-sheet and regulatory questions with citations, covering US/EU (TSCA, REACH) and Indian regulations. The agent loop is written directly on the Anthropic SDK, with Pydantic-validated tools and MCP integration. A real tool for my business, not just a demo.

Anthropic SDK FastMCP Pydantic RAG Multi-agent
classify→ dedup · cluster→ summarize→ faithfulness gate In progress

Meridian News

Self-hosted news summarization

A news aggregator on fully self-hosted models. Every ML stage starts from a cheap CPU baseline and only moves to a bigger model when it fails a stated bar on my own eval set. DeBERTa-v3 classifier with a calibrated "Other" class, embedding-based story clustering, and a pre-registered summarization bake-off scored on faithfulness, copy-rate and cost.

DeBERTa-v3 Embeddings FastAPI PostgreSQL MLflow
agent→ eval harness→ LoRA SFT→ GRPO Planned

Agent eval & distillation

Training a small tool-calling model

An evaluation harness for a tool-calling agent (LLM-as-judge with agreement checks, bootstrap significance tests, failure clustering), then used to train a small local model: LoRA fine-tuning first, then reinforcement learning with rewards verified against the mock API's state.

Unsloth GRPO LLM-as-judge Modal
crude · natgas · USD/INR→ model→ buy now / wait Planned

ChemSignal

Procurement timing signal

Tells a buyer whether to buy methanol now or wait, using public market drivers and scraped market prices. Evaluated walk-forward on rupees per tonne saved, and deployed on my own GitOps cluster rather than left in a notebook.

Time series Walk-forward eval GitOps
agents× ML engineering Exploring

Crucible

Agentic ML engineering

An agentic system for ML engineering work. Early exploration; more here soon.

Contact

Get in touch

The fastest way to reach me is email. I'm happy to talk about ML / LLM roles, or about any of the projects above.

Email me GitHub