Languages & APIs
- Python
- SQL
- FastAPI
- Pydantic
- REST API Design
- Async I/O
Open to new roles — Backend, AI/LLM & Platform Engineering
Sagar Arora
Backend Engineer AI & LLM Systems
I build production Python systems that hold up under real traffic — high-throughput FastAPI services, event-driven backends and LLM-powered RAG pipelines running on AWS and GCP. Currently shipping document AI tooling for a Warner Bros. Discovery team; previously the first backend hire at a US healthcare platform.
- Experience
- 4.3+ yrs
- Systems shipped
- 5+ prod
- Pipeline speedup
- −30% time
- LLM cost cut
- −50% spend
01 About
Backend foundations, AI on top
I'm a backend engineer who works close to the model layer. Over the last 4.3 years I've built and shipped production Python systems in two very different domains — US healthcare, where correctness and compliance are non-negotiable, and entertainment, where throughput and cost per unit of work decide everything.
At Collance I was the first engineering hire, which meant designing the backend architecture as well as the features: lab ordering with communication workflows, care coordination for social determinants of health, a call-centre platform on Telnyx, and a high-throughput power dialer.
At Elastiq I moved into AI systems work — entity extraction, sentiment analysis, region-wise social scraping, and RAG pipelines with contextual chunking and hybrid retrieval. That is where I learned to treat LLM spend and latency as engineering problems, not vendor problems.
Right now at Straive I work on DEX, a document AI utility platform for the Warner Bros. Discovery team — including a sensitive-entity masking module that detects and redacts PII before anything reaches a model.
I care about the unglamorous parts: query plans, index strategy, idempotency, structured logging, and pipelines that fail loudly instead of quietly.
02 Skills
What I work with
The stack I reach for most. Depth in Python and data systems; working fluency across the AI tooling I actually use in production.
Data & Storage
- MySQL
- PostgreSQL
- Redis
- BigQuery
- SQLModel
- SQLAlchemy
- Query Optimization
- Indexing
Cloud & Infrastructure
- AWS
- GCP
- Docker
- Kubernetes (GKE)
- Terraform
- EC2
- S3
- Lambda
- IAM
- CloudWatch
LLM Engineering
- LLM Integration
- Prompt Engineering
- Function Calling
- Tool Use
- LLM Evaluation
- Guardrails
- LangChain
- LangGraph
RAG & Retrieval
- Chunking Strategies
- Contextual Chunking
- Document Ingestion
- Vector Databases
- Semantic Search
- Hybrid Search
- BM25
- Reranking
- Query Rewriting
- Knowledge Bases
Quality & Delivery
- CI/CD Pipelines
- Grafana
- SonarQube
- Git
- GitHub
- Docker
- Linux
- Playwright
Daily tools
- PyCharm
- VS Code
- Postman
- DBeaver
- AWS CLI
- gcloud CLI
- Claude
- GitHub Copilot
- Google Gemini
- macOS
- Windows
03 Experience
Where I've built things
-
Senior Data Science Engineer
Straive (Gramener) Warner Bros. Discovery team
- Working with the WBD team on DEX, a document AI utility platform — contributing since project inception.
- Built a sensitive-entity masking module that detects and redacts PII and confidential data inside documents, strengthening data privacy and compliance.
- Designed and shipped an in-platform quick guide and guided tours feature that measurably improved onboarding and platform adoption.
- LLM
- Document AI
- PII Redaction
- Python
- Product
-
Software Engineer
Elastiq Technologies
- Optimised a Python entity extraction pipeline, cutting processing time by 30% and LLM cost by 50% through batching.
- Built an end-to-end sentiment analysis system, plus 5+ Python scrapers for region-wise social media insight.
- Implemented contextual chunking in a RAG data-parsing system, improving token efficiency and LLM output quality.
- Worked on a multi-layer prompt architecture handling routing, security and response synthesis in LLM workflows.
- Led onboarding for 3 enterprise clients, deploying end to end on AWS and GCP.
- LangChain
- RAG
- Scraping
- AWS
- GCP
- Prompt Design
-
Software Developer First engineering hire
Collance Technologies Pvt. Ltd.
- First engineering hire — built 5+ Python backend systems in the US healthcare domain.
- Developed LabPortal, a lab ordering platform with communication workflows and Five9 integration.
- Built the Care Coordination module for SDOH data collection, caregiver workflows and task management.
- Contributed to Outreach-IQ, a scalable call-centre platform integrating Telnyx.
- Implemented a high-throughput power dialer.
- Designed event-driven backend architecture for concurrent workflow execution at scale.
- FastAPI
- MySQL
- Redis
- Event-Driven
- Telnyx
- Five9
-
Cloud Support Associate — Big Data
Amazon Web Services (AWS)
- Developed Python scripts to parse logs and debug failures in EMR-based data processing jobs.
- Automated troubleshooting workflows, reducing manual effort in issue diagnosis.
- Built Python utilities for monitoring cluster performance and identifying bottlenecks.
- EMR
- Python
- Big Data
- Monitoring
04 Case Studies
Three problems, and how I solved them
Selected work, unpacked: the constraint, the approach, and what actually changed.
Straive · Warner Bros. Discovery
Sensitive-entity masking for a document AI platform
A compliance-critical redaction layer that runs before any document text reaches a model — so PII never leaves the boundary in the first place.
- Python
- LLM
- PII Detection
- Document Parsing
- Compliance
The problem
DEX ingests real production documents for a media client. Those documents carry personal and confidential data, and they were flowing downstream into AI processing largely untouched. Redaction that runs late — or misses edge cases — is a compliance problem, not just a data-quality one.
The approach
I built a masking module that sits in-platform and detects sensitive entity classes within document content, then redacts those spans in place. Keeping detection and redaction as distinct concerns means the detector can be tuned per data class without touching the redaction contract, and masking happens at the boundary so the raw text is never the version that gets handed on.
The impact
Strengthened the data privacy and compliance posture of the platform for a tier-one entertainment customer — and made document handling safe to expand to new, more sensitive content types.
Elastiq Technologies
Cutting LLM spend in half on an entity extraction pipeline
Latency and cost per record were the bottleneck on a region-wise social intelligence product. Batching plus a layered prompt architecture fixed both.
- −50%LLM cost
- −30%Processing time
- 5+Scrapers shipped
The problem
Entity extraction ran per-record against an LLM. The logic was sound, but every record paid full round-trip latency and full prompt cost — which made the pipeline expensive and slow exactly when volume went up. Regional coverage meant more scrapers feeding it, so the cost curve was getting steeper, not flatter.
The approach
The main lever was batching: group records so each model call carries more work, amortising latency and prompt overhead across a batch instead of paying it per record. Alongside that I worked on a multi-layer prompt architecture that separates concerns — routing, security policy, and response synthesis — so each layer could be tuned and evaluated independently rather than one monolithic prompt doing everything badly.
The impact
30% less processing time and 50% lower LLM cost, which made the region-wise product viable at volume. I also built the sentiment analysis system on top of it and 5+ Python scrapers feeding region-level social insight.
Elastiq Technologies
Contextual chunking that made a RAG pipeline cheaper and sharper
Naive fixed-size chunking was cutting documents in arbitrary places. Making chunks context-aware improved answer quality and cut wasted tokens.
- RAG
- Contextual Chunking
- Token Efficiency
- Hybrid Search
- BM25
- Reranking
The problem
The data-parsing stage split documents on raw token boundaries. A chunk could lose the heading or definition it depended on, so retrieval returned text that looked relevant and wasn't — and the fix-up prompting needed to compensate burned tokens on every single chunk.
The approach
I implemented contextual chunking: each chunk is enriched with the surrounding document context before it is embedded and indexed, so it carries the meaning it had in situ. The retrieval side follows the same philosophy — semantic search for meaning, BM25 for exact terms, combined as hybrid search and closed out with reranking so the best chunk wins, not just the first plausible one.
The impact
Measurably better token efficiency and LLM response quality. The same hybrid-retrieval thinking shows up again in the AI/LLM skill set — BM25, reranking and query rewriting are part of how I build retrieval, not separate buzzwords.
Personal Project · 2025
Uttaram.AI
An entertainment-domain platform that automates revenue calculation for movie distribution workflows.
- Python
- FastAPI
- PostgreSQL
- Redis
- BigQuery
- Docker
- Kubernetes
- Playwright
- LLM
- DigitalOcean
What it does
Revenue calculation and distribution efficiency for movie distribution workflows are manual, error-prone and slow. Uttaram.AI automates the pipeline and gives the business back the time it was losing to reconciliation.
Highlights
Built LLM-powered automation pipelines for the revenue side, and Playwright-based social media scrapers for region-level movie buzz detection. The buzz signal feeds directly into distribution decisions.
Impact
Contributed to a 30% revenue increase by improving distribution efficiency. Runs containerised on Kubernetes with Redis caching, PostgreSQL for state and BigQuery for analytics.
05 Education & Credentials
Background
Education
Master of Computer Applications
Amity University (Online)
Bachelor of Computer Applications
Shyam Lal College (IGNOU)
Certifications
- SQL Certification HackerRank
- AWS Cloud Practitioner Amazon Web Services
- Python Bootcamp Udemy
Recognition
- R&R Award Collance Technologies · 2025
- R&R Award Collance Technologies · 2024
06 Contact
Let's talk about your backend
Hiring for backend, AI/LLM or platform roles — or just want to compare notes on a RAG pipeline? My inbox is open.