AI engineer · Oracle · Bengaluru, India

Aprameya
C V

I build agentic and generative AI systems, and measure what they actually do.

I take AI from a customer's first conversation to production at Oracle: retrieval, agent design, GPU serving, and the infrastructure underneath. Every project on this site ships with its own benchmark and its own list of what it does not prove.

Role
Cloud Engineering, AI
At
Oracle, EMEA customers
Open to
Applied AI, forward-deployed
Scroll

About

Who I am

An engineer who likes questions with a measurable answer.

I am an AI engineer at Oracle in Bengaluru. I joined as a project intern in January 2025, moved into a full-time engineering role that summer, and now work as a technical scoping lead for customers across EMEA.

Most of my day is the gap between a prototype and something a customer can rely on: scoping the problem in the first conversation, building the agents, retrieval and serving layers, and deploying them on OCI. I have built systems for government and tax-authority clients, including multilingual assistants and image-matching APIs that run on a CPU-only machine.

Outside client work I run small, controlled studies on the same questions: how many passages an agent should retrieve, which serving flag actually helps, when a cheap model should hand off to a strong one. Each is published here with its numbers and its limits.

I studied electronics at JSS Science and Technology University, where I published on SLAM and hold a design patent. That is where the habit comes from: state the question, fix the variables, report the interval.

0xLLM serving throughput from speculative decodingSpecLane
0%right clause found, up from 47%, in half the timeHBC-RAG
0%critical defects caught at 47% of the strong model's costCG-BI
0 ptsaverage gap between stated confidence and accuracyPramana

03 / Method

How I measure

The same four habits run through every project on this site.

01

Change one thing at a time

Each study fixes the model, the prompt and the data, then varies a single setting. HBC-RAG varies passages per hop. SpecLane varies one engine flag. A difference then has one possible cause.

02

Score without a judge

Character spans, exact match and held-out quality guards, not an LLM grader. Every number regenerates from saved runs, with no API spend and no judge drift.

03

Intervals, not single runs

Comparisons carry paired bootstrap intervals. Anything inside the interval is called noise, including the cases where my own result is not significant.

04

State what it does not claim

Every project page ends with its limits: one seed, small samples, no healthy class, one document per question. A result is only useful with its boundary.

04 / Experience

Oracle

Three roles in under two years, from intern to technical scoping lead.

Jun 2026 to present

Oracle / Cloud Engineering, AI (L2)

  • Technical scoping lead and point of contact on EMEA engagements across GPU serving, guardrails, voice AI, MCP and agentic automation.
  • Built a three-agent ServiceNow suite for a public-sector client: natural-language search over 1M+ tickets in under 7 seconds, hybrid search over 500K+ multilingual articles, and ticket creation in under a minute.
  • Built a multilingual legal and tax RAG assistant for a national tax authority, so citizens can ask questions in their own language.
  • Delivered OCI infrastructure end to end: landing zones, networking, storage, IAM, VM migration and disaster recovery, automated with Terraform.
  • Built deployable solution packs for RAG, agents, voice, MCP, guardrails and agent memory, and drove Codex adoption for 80+ people who built 40 to 50 reusable assets.
Jul 2025 to May 2026

Oracle / Cloud Solution Engineer (L1)

  • Built a trademark platform for a government ministry, exposed as two APIs: similarity over 500K+ images with multilingual text and compliance rules on a CPU-only machine, and open-source image generation and editing on NVIDIA CUDA GPUs through a custom ComfyUI module.
  • Primary developer of internal AI tools used across teams, from a document generator to a security-review agent, saving an estimated 1,100+ hours.
Jan 2025 to Jul 2025

Oracle / Project Intern

  • Built the Solution Document Generator on LangGraph and OKE, contributing to 50+ documents and 100+ hours saved.

05 / Stack

What I work with

Grouped by the layer it sits in, not by logo.

Agents and retrieval

ADK, LangGraph, LangChain, Deep Agents, OpenAI SDK, MCP, hybrid retrieval, pgvector, guardrails, agent memory

Inference and training

vLLM, llama.cpp, speculative decoding, FP8 KV cache, CUDA, NVIDIA GPUs, LoRA fine-tuning, Hugging Face

Cloud

OCI (Enterprise AI, Dedicated AI Clusters, Container Instances, OKE), AWS (EC2, ECS, Lambda), Google Cloud (Vertex AI, Model Garden)

Infrastructure

Landing zones, networking, storage, IAM, VM migration, disaster recovery, Terraform, Kubernetes, Helm, K3s, Prometheus

Languages and tooling

Python, TypeScript, SQL, FastAPI, Linux, Docker, CI/CD

06 / Education

Research habit

JSS Science and Technology University / B.E., Electronics and Communication

Dec 2021 to Jun 2025 · CGPA 8.93 · Gold Medal

Published on SLAM and VSLAM in Springer and IEEE venues and hold a design patent. I kept the research habit: state the question, fix the variables, report the interval.

07 / Contact

Building something that needs to work?

I am open to applied AI and forward-deployed engineering roles. Email is the fastest way to reach me.

Email copied