Akash Ungarala
Senior Software Engineer
I build resilient, high-scale backend systems — and the production AI systems that run on them. 10+ years architecting real-time platforms processing billions of events across broadcast media, fintech, gaming, ad tech, and energy.
About
Systems that scale when it matters most
10+ years building backend systems where failure isn't an option. Real-time broadcast captioning for 8M+ viewers. Ad platforms processing 2B+ daily events. Observability tools trusted by 50+ engineering teams. I design systems where uptime and throughput are non-negotiable.
Great backend systems are invisible: fast, consistent, resilient under load. I combine Python's expressiveness with TypeScript's type safety to build services teams can maintain and scale. I own problems end-to-end — architecture, deployment, and on-call.
Most recently built NGMM at Marathon Petroleum — a multi-agent AI assistant that grounds LLM reasoning in refinery maintenance data, with the Azure AD auth, session persistence, and OpenTelemetry tracing it takes to survive production. Same discipline as the broadcast and fintech work, pointed at a newer problem.
Core Expertise
Tech Stack
Languages & Frameworks
Distributed Systems
Cloud & Infrastructure
Platform & DevOps
AI & Agents
Data & Observability
Background
Experience & Education
Senior Software Engineer, AI Platform
Marathon Petroleum Corporation (Contract) · Remote
Feb 2026 — Present
Built two enterprise AI systems: NGMM, a multi-agent assistant grounding LLM reasoning in refinery maintenance data, and Mara-Terra, an MCP server exposing internal infrastructure knowledge to AI clients.
- Architected multi-agent orchestration in Python/FastAPI with Semantic Kernel — SQL retrieval, planner, graph, and notification agents under a coordinating manager — reducing manual investigation time by 30-40% and analyst research effort by 50%
- Engineered Cosmos DB session persistence with turn reconstruction and tool-call rehydration, improving continuity by 90% across long-running investigations; secured access with Azure AD, JWT validation, and RBAC
- Instrumented agent execution end-to-end with OpenTelemetry and Azure Monitor, adding span propagation and a usage/trace dashboard that cut issue diagnosis time by 35%
- Built Mara-Terra from scratch — an MCP server in Python and FastMCP exposing Terraform and pipeline module search to AI clients over Azure AI Search, with read-only, idempotent tool contracts and managed-identity auth
- Hardened both platforms with CI quality gates, dependency pinning, and lint enforcement, reducing regressions by 40% and doubling release velocity
Architected CaptionExchange, a real-time AI/human captioning platform serving 8M+ concurrent viewers with sub-5-second failover for live broadcast operations.
- Designed fault-tolerant captioning orchestration layer reducing failover from 3 minutes to <5 seconds through predictive routing and automated provider switching
- Architected high-concurrency WebSocket gateway handling 10K+ simultaneous connections with sub-100ms p99 latency for real-time audio and caption exchange
- Led technical design reviews and mentored 2 engineers on distributed systems patterns, improving team velocity by 25%
- Established CI/CD with Semgrep security scanning and 90%+ test coverage, reducing production incidents by 60%
Designed centralized SLO tracking platform processing metrics from 200+ services, reducing Mean Time to Detection by 40% across 50+ engineering teams.
- Built real-time observability dashboard integrating OpenTelemetry traces from 500+ microservices with automated breach detection via SNS
- Led proof-of-concept for high-cardinality metrics using Amazon Managed Prometheus, reducing alert noise by 50% and saving 20 engineering hours/week
- Drove Terraform standardization across 12 observability services, reducing deployment failures by 70% through enforced CI/CD gates
- Owned full stack: FastAPI backend, React frontend, PostgreSQL, AWS Lambda infrastructure with Terraform
Owned AdTech backend platform for Pokemon GO serving 80M+ monthly users with 99.99% uptime, processing 2B+ daily events for location-based ads.
- Designed event-driven pipeline using Celery and Cloud Spanner, reducing p99 latency from 800ms to 200ms for in-game reward delivery
- Optimized Spanner query patterns and indexing strategy, cutting analytics query latency by 60% for 50M+ daily events
- Reduced production defects by 40% through TDD adoption and hardened CI/CD gatekeeping; mentored junior engineers on testing strategies
- Built WebAR ad telemetry and tiered reward systems powering contextual ads for millions of players globally
Re-architected identity resolution services scaling 4x (5K to 20K+ writes/sec) while reducing infrastructure costs by 30%.
- Engineered event-driven data pipelines processing 5TB+ daily via BigQuery and PostgreSQL, powering 200+ enterprise partner integrations
- Reduced query latency by 30% through database optimization and implemented RBAC security using Auth0 for partner API access
- Orchestrated complex ETL workflows using Apache Airflow DAGs with automated deployments via Spinnaker
- Designed Flask-based microservice templates adopted company-wide, reducing service bootstrap time from 2 weeks to 2 days
Built Dialogflow conversational AI prototypes deployed across 400M+ Google Assistant devices.
- Developed chatbot prototypes using Python and Node.js, collaborating with product teams to define integration patterns for enterprise customers
- Implemented webhooks processing natural language requests with real-time dynamic responses
Built enterprise Django portal with MySQL backend serving 10K+ users for PG&E (Pacific Gas and Electric Company).
- Developed user management modules and CRUD operations with efficient SQL triggers and stored procedures
- First production engineering role — foundation in full-stack development, database design, and API integration
Portfolio
Featured Projects
Contact
Let's build something together
Looking for senior and staff-level roles where I can architect systems at scale. Real-time platforms, distributed systems, high-throughput pipelines — I own hard problems end-to-end.
Send a Message
Current availability
Remote (US) or Hybrid (NC, SC)