AI Architect

Revanth Gollapudi

I build AI systems that run in production.

I architect, build, and own the AI systems that run inside large companies — RAG platforms, evaluation infrastructure, model-routing pipelines, and the deployment plumbing that keeps them up at 99.9%.

Headline numbers
85K
requests / day
6,000
users served
99.9%
uptime
−30%
inference cost
+0.12
RAGAS faithfulness
Experience

Where I've been working

AI Architect · CGI

Dec 2024 — Present · Remote

Where I am now. The work splits roughly into three buckets: building the platforms that host AI systems, making sure those systems give correct answers, and keeping the costs sane.

HR Assistant · 6,000 employees

Sole architect on an end-to-end Azure RAG system for an HR Policy chatbot. Hybrid retrieval (vector + BM25 + semantic reranker), GPT-4o on PTU, Document Intelligence, Entra ID/MFA, private VNet. <3s response, ~40% cache hit rate, 99.9% uptime, EU AI Act-aligned.

Conversational AI Platform · 85K req/day

Architected and own a Kubernetes-based platform on AWS EKS serving ~85K requests/day across 10+ agent workflows for 1,200 users. Lead a 5-person team with full SLA accountability.

Inference cost · −30%

Co-designed a complexity-based model router sending simple queries to Gemma/Mistral 8B and complex ones to GPT-4o. Cut monthly inference spend from $15K to $10.5K with no quality degradation.

Hallucination · 45% → 30%

Fine-tuned Llama 3 via LoRA/PEFT with DPO on 500+ human-labeled preference pairs. Reduced measured hallucination rate from 45% to 30% across production workloads.

EvalOps · regression gate in CI

Built a full EvalOps pipeline — groundedness rubrics, golden datasets, automated RAGAS scoring, OpenTelemetry retrieval tracing — as a mandatory CI gate before every model update. No quality regressions have reached production undetected since rollout.

AI Vehicle Inspection

Lead engineer on an AI-powered vehicle inspection system for an enterprise logistics client, replacing manual walk-around inspections. Drove the strategic pivot from a YOLOv8 model to GPT-4o Vision for zero-shot damage understanding.

AWS EKSAzureGPT-4o PTUGPT-4o VisionYOLOv8OpenCVAzure FunctionsCosmos DBLlama 3LoRA/PEFTDPORAGASOpenTelemetryLangGraphBicep

Data Engineer · CGI

Jan 2023 — Dec 2024 · Remote

Two years building the foundations: training pipelines, retrieval systems, evaluation tools — the infrastructure my later architecture work would rely on.

Retrieval faithfulness · 0.62 → 0.74

Tuned a hybrid retrieval system (BM25 + vector + cross-encoder reranking) with query rewriting. Lifted RAGAS faithfulness from 0.62 to 0.74 on a 500-question eval set — a +0.12 absolute jump on a hard-to-move metric.

20% of bad updates blocked

Built a CI/CD pipeline using LLM-as-judge with rubric scoring as a mandatory release gate. Caught about a fifth of regressive model updates before they reached users.

ML pipeline cycle · 72h → 24h

Rebuilt Kubernetes ML training pipelines for a 6TB document classification workload. Cut the train-to-deploy cycle from three days to under one — a 3× faster iteration loop.

Observability · hours to minutes MTTD

Implemented CloudWatch and OpenTelemetry instrumentation with drift detection and alerting. Reduced mean time to diagnose failures from hours to minutes.

BM25pgvectorCross-encoderRAGASLLM-as-judgeKubernetesCloudWatchOpenTelemetry

Data Engineer · eKincare

Aug 2020 — Jul 2021 · Hyderabad, India

My first proper engineering job — at a healthcare startup. Mostly streaming data pipelines, dead-letter queues, and learning what production really means.

+40% pipeline throughput

Tuned partitioning and parallelism on the Kafka and Spark setup. Added schema validation and dead-letter handling so a single bad record couldn't take down the whole pipeline.

−15% application crashes

Tracked down recurring failure patterns and rebuilt the logging, monitoring, and alerting.

Apache KafkaApache SparkPythonAWS
Live demo

What I actually do, made visible

A lot of my work is invisible — it's the part of an AI system that happens between you typing a question and the answer appearing. Below is a small interactive version, using my own work history as the “knowledge.”

Type a question (or pick a preset), and you'll see the steps the system goes through: breaking down the question, searching for information, ranking what it found, checking the answer is trustworthy, and writing a response.

presets:
>
Side projects

Things I built on my own time

Side projects where I get to try out ideas I've been thinking about. All open source.

RAOC

↗ GitHub
Remote control your Mac from your phone

I built this because I kept needing to grab files from my home Mac while I was out. RAOC lets you send a natural-language command from Telegram and a six-agent pipeline figures out the plan, asks you to approve it, then runs it. The interesting design problem was safety: every plan needs tap-to-approve, dangerous commands are blocked at the code level, and every file is backed up before being touched. 303 tests in CI to keep it that way.

Python 3.12Claude APITelegramSQLitePydantic v2

LLM Eval Harness

↗ GitHub
A serious framework for testing AI changes before you ship them

Most teams test their AI by eyeballing examples — that doesn't catch regressions. This framework does it properly: deterministic checks first (format, citations), then an LLM-as-judge with rubric scoring, then 95% bootstrap confidence intervals on the win rate. If quality drops below threshold, the build fails with exit code 2.

PythonOpenAI APIRAGASGitHub Actionspytest

FIFA World Cup 2026 Forecasting

↗ GitHub
Calibrated tournament simulation for the 2026 World Cup

A calibrated probability model for the 2026 World Cup. A bivariate Poisson core with Dixon-Coles correction learns expected goals; an LLM layer reads news and lineups and emits only typed categorical signals, never numbers; a Monte Carlo simulator runs 50,000 tournaments for stage-by-stage odds. The betting market is used only to grade the model, never to train it, enforced by a leakage test.

PythonNumPyMonte CarloDixon-ColesClaude APIFastAPIpytest
Education

Where I studied

M.S. Applied Engineering

University of Colorado Boulder · 2021 — 2022

B.Tech. Mechanical Engineering

PES University · 2016 — 2020
Get in touch

Let's build something

Senior or lead AI engineering roles at teams that are serious about reliability, cost, and measurable quality — not just shipping demos. Especially interested in companies building AI products where the stakes are real: enterprise tools, healthcare, finance, infrastructure.

The fastest way to reach me is email. I read everything — so if you're building something interesting, send a note.

gollapudi.off@gmail.comLinkedInGitHub