Shree Bohara

Shree Bohara

Software Engineer @ QuinStreet

I build production AI systems, and the infrastructure for not trusting them

About Me

I'm a Software Engineer at QuinStreet in San Francisco, working on Pond (insurance.com/pond). I joined as an intern in June 2025, helped take Pond from an empty repository to production in under two months, and converted to full-time in June 2026. Most of my work is applied AI in production — a conversational insurance advisor built end to end, an incident pipeline that reads live production error streams and explains them — plus the layer underneath that refuses to trust a model: containment for AI agent swarms, an auditor for reported benchmark gains, and fail-closed gates in a trading system that has deliberately never placed an order. When it helps, I go a level deeper than most application engineers: an LLM inference engine written from scratch in C++17, and a cache-blocked Bloom filter inside DuckDB's hash join.

Skills & Technologies

Languages

PythonJavaC++TypeScriptJavaScriptSQLBashHTML/CSS

AI & LLM

Anthropic Claude APIOpenAI APIRAGVector SearchTool CallingStructured OutputsPrompt EngineeringLLM EvaluationTree-sitterSentence-Transformersscikit-learn

Backend

Spring BootSpring WebFluxFastAPIFlaskNode.jsExpressRESTSSEWebSocketsWebRTC

Frontend

ReactNext.jsVue 3Tailwind CSSZustandReact FlowThree.jsreact-three-fiber

Data

PostgreSQLMySQLMariaDBMongoDBRedisElasticsearchChromaDBSQLiteDuckDBBigQuery

Cloud, DevOps & Observability

AWSGCPDockerAnsibleCI/CDVercelRenderELK StackKibanaStructured LoggingDistributed Tracing

Testing

PyTestJUnitVitestPlaywrightMutation Testing

Experience

QuinStreet

Software Engineer

QuinStreet

Current

Jun 2026 - Present · San Francisco, CA

  • Built and operate an AI incident pipeline — alerting rules over Elasticsearch, a FastAPI service that fingerprints and batches errors into incidents, Claude analysis grounded in the actual source files, and cards delivered to the team channel. It runs on dev and stage and watches production error streams; several hundred raw events collapse into a few dozen incident reports a day, each arriving with the log link and the analysis already done(Root cause correct on 48 of the last 50 reports I scored by hand)
  • Cut the pipeline's projected LLM spend with a per-fingerprint analysis cache, selective prompt caching, delivery-gated spend and quota-aware backoff, while raising the context budget rather than lowering it(Projected 70–85% reduction)
  • Designed and shipped the consent re-capture flow for a consumer rebrand: a consent state machine instead of a boolean, every capture path funnelling into one idempotent endpoint with a single-winner atomic transition, the audit row written in the same transaction, and exactly-once propagation downstream. Live in production since July 2026
QuinStreet

Software Engineer Intern

QuinStreet

Jun 2025 - Jun 2026 · San Francisco, CA

  • Co-architected Pond on AWS and helped take it from nothing to production in under two months, then built the multi-step onboarding most accounts are created through: address autocomplete and normalisation, cascading make/model/trim selection behind staged cache lifetimes, and policy import that parses an existing policy instead of making people retype it
  • Built the rating engine in Java and Spring Boot over carrier rate APIs — the piece that turned Pond from a place to organise your insurance into a place to get quoted
  • Built Ollie, Pond's conversational insurance advisor, end to end: nine routed conversation contexts, a Spring WebFlux backend streaming over SSE, a React client that renders partial markdown correctly mid-stream, and in-chat coverage changes instead of a hand-off to a multi-screen form(A short-TTL dedup cache cut redundant inference calls by roughly 40%)
DeepTek Medical Imaging

Software Engineer Intern

DeepTek Medical Imaging

Jun 2023 - Jan 2024 · Mumbai, India

  • Built the FHIR-compliant Health Identifier platform in Node.js, React and MySQL that moved patient records between hospitals in a single standard shape(500+ hospitals, 100K+ monthly users, 50K+ record exchanges a day)
  • Shipped OAuth 2.0 authentication, role-based access control and PostgreSQL audit logging; the platform passed the ABDM compliance review on first submission
  • Cut API response time 3× with Redis caching on the hot paths and held 99.9% uptime across the hospital integrations, with distributed tracing and structured logging across six services

Education

University of Southern California

University of Southern California

Master of Science in Computer Science

2024 - 2026 · Los Angeles, CA

Machine LearningDatabase SystemsInformation RetrievalAnalysis of AlgorithmsWeb Technologies

Let's Connect

I'm always open to a conversation about AI infrastructure, agent systems, or production reliability.