Subburajan Perumal

Full Stack AI Engineer — LLM Agents & Cloud Systems

I build AI-assisted logistics tooling, cloud ETL pipelines, and full-stack applications. Google Cloud Associate Cloud Engineer and AWS Certified Cloud Practitioner. Most of my recent work is in Python, FastAPI, React, and LLM agent frameworks — usually solving a specific data or automation problem rather than shipping a demo.

Chennai, Tamil Nadu, India

Get In Touch View Projects Resume

Experience

Engineering career timeline

Software Engineer

Freightify

Jan 2026 – Sep 2026· 9 months

Chennai, Tamil Nadu, India · On-site

–Engineered an autonomous AI agent on Gemini and Claude (Gemini for the low-cost first pass over PDFs with images, Claude for the steps after it) to extract ocean freight rates from major carriers including MSC, Maersk and ONE, from unstructured Excel and PDF documents, for a platform whose clients include DP World
–Built a serverless recipe-driven ETL pipeline on Google Cloud Run Jobs, ingesting carrier rate workbooks from GCS and normalizing them into Next-Generation Tariff (NGT) schemas — reducing processing turnaround by 80%
–Built a Go backend for search and rate-matrix standardization, powering the downstream rate-ingestion matrix
–Implemented sliding-window token optimization and deterministic business rule execution to eliminate LLM hallucinations in financial rate data

PythonGoGemini APIClaude APIAI AgentsGCPCloud Run JobsKubernetesDockerETL

Senior Software Engineer

Door Sabha Nigam Limited

Aug 2025 – Dec 2025· 5 months

Chennai, Tamil Nadu, India

–Promoted from Software Developer in recognition of contributions across the conference portal, voice messaging platform, and AWS migration work below

PythonFastAPIAWSReactWebSocketsAlgolia

Software Developer

Door Sabha Nigam Limited

Jul 2023 – Jul 2025· 2 yrs 1 month

Chennai, Tamil Nadu, India

–Built a conference API portal for ICICI and BFIL, and a voice messaging platform for Reliance and BFIL
–Contributed to a MultiCall Android application published on the Google Play Store, reaching 30,000+ active users
–Led cloud migration (on-premises → AWS) achieving 40% infrastructure cost reduction
–Integrated the Algolia AI search engine into the platform
–Managed the reporting process and built an ETL pipeline to sync data across internal systems

PythonFastAPIAWSReactWebSocketsAlgolia

AWS re/Start Graduate

Amazon Web Services (AWS)

Feb 2023 – Apr 2023· 3 months

Chennai, Tamil Nadu, India

–Completed intensive AWS cloud training program covering EC2, S3, RDS, Lambda, IAM, and networking
–Earned AWS Certified Cloud Practitioner certification upon program completion
–Built hands-on lab projects deploying secure, scalable cloud architectures

AWSCloud ArchitectureNetworkingSecurity

Automation Intern

Ghostengines Technologies Pvt. Ltd.

Dec 2021 – Jun 2022· 7 months

Chennai, Tamil Nadu, India

–Automated lead scoring, routing, and duplicate detection workflows using Python and RPA tools
–Achieved 70% reduction in data entry errors and 40% decrease in lead processing time
–Reduced manual operational overhead by ~80% across internal sales teams

PythonRPAProcess Automation

Projects

RAG, concurrent ingestion pipelines, and workflow automation — each with its code

AI & Document Intelligence

2026

Logistics Rate RAG — Deterministic Retrieval-Augmented Q&A

A public, clean-room RAG agent over freight-rate and tariff documents, built to prove an LLM answer can be made deterministic and safe, not just fluent. Hybrid dense + BM25 retrieval, fused with reciprocal rank fusion, feeds a cross-encoder reranker — then every answer must clear a schema-parse gate, a business-rule gate, and a grounding/confidence gate before it's shown, or it's refused instead of guessed.

–Hybrid dense (Chroma / Pinecone) + BM25 retrieval fused with RRF, reranked with a cross-encoder before generation
–Deterministic guardrail layer — schema-enforced output, business-rule validation, grounding/confidence gate — with zero hallucinated rate values on the adversarial test set
–Live demo answers real questions with the exact cited source chunks; the full gate trace lands as the eval harness completes

Code Live Demo

LangChainRAGGemini APIChromaPineconePython

Cloud & Distributed ETL

2026

Concurrent Go Ingestion → Kubernetes → PySpark Rate Pipeline

A public, clean-room rebuild of a production rate-ingestion pipeline: a Go worker-pool service pulls multiple carrier feeds concurrently, runs containerized on Kubernetes (Helm, kind + EKS), and hands off to a PySpark job that normalizes and loads the result into Redshift — one real dataset proving Go concurrency, Kubernetes operations, and PySpark end to end.

–Go worker pool (goroutines, channels, context cancellation) ingests 8 synthetic carrier feeds concurrently — 20,810 records/run, 1.82× measured speedup, race-clean under go test -race and goleak
–Deployed via Helm to kind and EKS with liveness/readiness probes and a graceful-shutdown e2e test — pod terminates within 1s of a 60s grace with no partial writes
–PySpark job normalizes 20,810 → 2,294 rows (18,450 dedup, 66 named rejects) and loads to Redshift; CI runs go test, kind e2e, and PySpark tests, all green

Code Results Live Demo

GoKubernetesHelmPySparkRedshiftDocker

Automation & RPA

Dec 2021 – Jul 2023

Automated Lead Scoring & Routing System

Lead automation for a real-estate platform: property-enquiry leads from Zoho CRM, MagicBricks, and 99acres are matched to builder projects by keyword and registered on each matching builder's website by queued Selenium workers. Screenshot proof is synced back to the CRM. This cut errors 70%, processing time 40%, and manual overhead ~80%.

–Flask webhook + Celery/Redis with four routed queues (lead, browser, bulk, portal sync), so bulk CSV uploads never delay live leads
–25 per-builder Selenium form-fillers with a 30-day dedup window and SUCCESS / PARTIAL / FAILED status tracked per lead in MongoDB
–Scheduled jobs: hourly MagicBricks + 99acres sync, Zoho OAuth token refresh cached in Redis, and screenshot retention cleanup
–70% error reduction, 40% faster lead processing, ~80% less manual overhead

Code

PythonFlaskCeleryRedisMongoDBSeleniumZoho CRMDocker