01 / engineering portfolio

PraveenMandala

AI Engineer & Backend Engineer

I build dependable AI and backend systems—from grounded RAG pipelines to event-driven infrastructure that is built to be measured, operated, and improved.

View projects Resume

FocusAI infrastructure & backend systems

Based inAndhra Pradesh, India

Scroll to exploreSelected systems

02 / ContextThe person behind the systems

Computer Science student focused on Generative AI, backend engineering, and distributed systems. I like working where model behavior, data systems, and production constraints meet.

My work spans retrieval-augmented generation, LLM evaluation, event-driven architectures, authentication systems, and machine learning infrastructure. I care about clear boundaries, measurable behavior, and systems that stay understandable as they grow.

Education: B.Tech in Computer Science & Engineering · 2023–2027
Aditya College of Engineering & Technology · CGPA 8.65 / 10
Working on: Generative AI, backend engineering, distributed systems
Contact: praveenm0088@gmail.com

03 / Capability index

Tools for serious systems.

A working map of the languages, services, and infrastructure behind the projects.

Grounded generation, evaluation, and model-driven product systems.

Generative AILLMsRAGNLPPrompt EngineeringAgentic AILangChainLangGraphHugging FacePyTorchVector SearchpgvectorOpenAI APIGemini API

04 / Selected systems

Built to be explored.

Four systems, four different constraints. Scroll through the architecture, decisions, and outcomes behind each one.

RAG customer support platform

QUERY / SUPPORT TICKET

SEMANTIC

HYBRID RETRIEVAL

BM25

GROUNDED ANSWER 88% faithful

01 / RAG customer support platform

IntelliSupport

Enterprise support answers need to be grounded in internal knowledge instead of generated from model memory alone.

Engineering decisions

Use hybrid retrieval to protect recall when terminology is domain-specific.
Measure faithfulness and relevance separately so retrieval and generation regressions stay visible.

Impact

85% retrieval hit rate90% intent accuracy88% faithfulness · 82% relevance

GitHub

01 IntelliSupport 02 LLM Evaluation Framework 03 Real-Time Feature Store 04 NLP-Service

01 / chapter

RAG customer support platform

QUERY / SUPPORT TICKET

SEMANTIC

HYBRID RETRIEVAL

BM25

GROUNDED ANSWER 88% faithful

01 / RAG customer support platform

IntelliSupport

Enterprise support answers need to be grounded in internal knowledge instead of generated from model memory alone.

Engineering decisions

Use hybrid retrieval to protect recall when terminology is domain-specific.
Measure faithfulness and relevance separately so retrieval and generation regressions stay visible.

Impact

85% retrieval hit rate90% intent accuracy88% faithfulness · 82% relevance

GitHub

02 / chapter

LLM benchmarking platform

MODEL BENCHMARKv.04 / CI

.88

.91

.76

.83

.94

.79

.86

.72

.89

.81

.92

.84

.78

.95

.87

faithfulnessrelevancybertscore

02 / LLM benchmarking platform

LLM Evaluation Framework

RAG and LLM applications need repeatable evaluation before model, prompt, or retrieval changes reach users.

Engineering decisions

Combine model-based judgment with lexical and embedding metrics instead of trusting one score.
Run evaluation in CI so prompt changes are reviewed like code changes.

Impact

Faithfulness, context relevancy, and answer relevancyBLEU, ROUGE, and BERTScore coverageContinuous evaluation in the delivery workflow

GitHub Live system

03 / chapter

MLOps infrastructure

EVENT STREAM / LIVE

REDISonline serving12 ms

POSTGRESQLdurable historysource of truth

03 / MLOps infrastructure

Real-Time Feature Store

Real-time models need fresh features with low-latency reads while retaining durable historical data for analysis and recovery.

Engineering decisions

Split online and historical storage so serving latency does not depend on analytical queries.
Make writes idempotent and keep a cache-aside fallback for degraded reads.

Impact

Dual-storage serving modelLow-latency online feature accessContainerized event-driven deployment

GitHub Live system

04 / chapter

Asynchronous NLP pipeline

QUEUERabbitMQ

CELERY
worker 01

NLP
inference

CELERY
worker 02

SENTIMENT

NER

PERSIST

04 / Asynchronous NLP pipeline

NLP-Service

Model inference should not block HTTP requests when NLP jobs can be processed asynchronously in the background.

Engineering decisions

Move inference behind a queue so HTTP latency is independent of model runtime.
Persist job state so clients can inspect results without holding an open request.

Impact

Background sentiment analysisNamed entity recognition pipelineQueue-based, containerized inference

GitHub

05 / Experience

Shipping with a team.

Production experience grounded in security, API design, and the software delivery lifecycle.

Full Stack Developer Intern

Zengen Technologies

May 2025 – Aug 2025

Built secure authentication and authorization services and connected them to business modules in an Agile product team.

Designed OAuth 2.0 and JWT authentication APIs with refresh-token rotation and role-based access control.
Implemented bcrypt hashing, token expiry, rate limiting, request validation, and JWT verification middleware.
Collaborated across feature development, testing, debugging, and deployment workflows.

Node.jsExpress.jsOAuth 2.0JWTREST APIs

06 / Proof rail