AI Engineer · Production Systems · Delhi NCR, India
I’m Abhishek. I build the systems behind modern AI products—from inference and evaluation to context engineering and backend architecture—so they remain reliable in production.
10+ years building production software · Serving users in 85+ countries
Independent research · 2026
I tested whether a small recurrent model can solve harder problems just by thinking for longer. Training across a range of reasoning depths mattered far more than architecture. Auditing my own graph generator then undid the most exciting-looking result.
I enjoy building software that survives real users.
I joined AssessPrep as one of its first engineers in 2016. Over nine years, I moved from frontend to backend, infrastructure, and AI while helping the platform grow to serve users in more than 85 countries.
Today I spend most of my time improving the reliability, latency, and evaluation of production AI systems. I’m particularly interested in context engineering, inference efficiency, and the tooling that helps teams ship AI safely. Outside the day-to-day work, I’m usually testing an inference idea or reading AI systems research.
Production AI · AssessPrep
Developed and deployed fine-tuned models for automated assessment grading and content recommendations, including evaluation, lifecycle monitoring, and the MLOps pipelines needed to update them safely. Also built NLP systems for question analysis and an in-house RAG support bot.
Infrastructure · AssessPrep
Helped scale the platform to serve users in 85+ countries. Moved a monolith to Kubernetes-based services, automated delivery with Terraform and CI/CD, and introduced monitoring and autoscaling. Deployment time fell 70%, infrastructure cost fell 40%, and reliability reached 99.9%.
Working with Go · Python · Node.js · Kubernetes · Terraform · AWS/GCP
Good AI products are reliable before they’re clever, measured before they’re optimized, and useful before they’re impressive.
I ran controlled experiments to figure out why code LLMs hallucinate cross-file references. The answer: they lack navigation, not comprehension. 14 tokens of symbol names gets 76% of the way. Full files get 92%. And you can inject the signal without any text at all.
Read →A browser-based video editor that runs entirely client-side using FFmpeg compiled to WebAssembly — no uploads, no servers, full privacy.
Investing in Prometheus and Grafana did more than fix outages — it revealed performance bottlenecks, improved reliability, and changed how we think about building software.