Keshav Malik | AI Engineer

Let's Talk [t]work [w]resume [r]github [g]

Hi, I'm Keshav Malik

AI Engineer · LLM Systems & Production Deployments

Available for work [t]

I build AI systems and take them into hard real-world environments. LLM extraction and retrieval, multilingual speech pipelines, evaluation against human-verified ground truth, and the deployment work that gets all of it running across cloud, on-prem and fully air-gapped networks where nothing can leave the client's environment. Currently forward-deployed across India's financial sector.

Before that, supply chain automation and full-stack product work for startups in the US and India.SIH’24 Winner, ex-GDSC Lead, and someone who would rather ship something into production than talk about it.

~ What I Work On

Production AI Systems

Speech and LLM pipelines running against real enterprise data, not demos.

›Multilingual speech-to-text pipeline processing 70+ hours of audio a day across AWS, Azure and GCP
›LLM extraction with retrieval over client entity masters, structured outputs, and per-asset-class schema validation
›An LLM adjudication layer that reconciles extracted deals against trade files, with a confidence bar and an explicit path to human review
›Evaluation against human-verified gold sets, scored on task outcomes rather than proxy metrics

Python

LiteLLM

vLLM

Bedrock

Langfuse

Prefect

Deployments Into Hard Environments

One product, many institutions, four substrates, including networks with no internet access.

›Shipped to asset managers, brokerages, primary dealers, a national exchange and the market regulator
›One image everywhere, behaviour selected by environment, so there are no per-client builds to maintain
›Air-gapped delivery through signed image tarballs with checksums and manifests, rolled forward by an updater on machines I cannot reach
›Cloud on EKS with Helm and ArgoCD; on-prem across AWS, GCP and bare metal with no cloud services at all
›Per-client identity: Entra ID SSO, LDAP/AD, and RSA-encrypted credential flows
›Observability that reports latency, tokens and cost while masking content, so nothing sensitive leaves the client network

Docker

Kubernetes

Helm

ArgoCD

AWS

GCP

Azure

Bare metal

Full-Stack Product

The application layer that people actually touch.

›Next.js and React front ends over FastAPI, Flask and Express services
›MongoDB, PostgreSQL and Redis, with Prefect for pipeline orchestration
›Shipped products across financial compliance, supply chain automation, real estate and healthcare

TypeScript

Next.js

React

FastAPI

Node.js

MongoDB

PostgreSQL

~ Work Experience