Software Engineer | Data Scientist | ML Engineer | AI Engineer
Software Engineer specializing in Data Science Infrastructure and Distributed Systems, with 3+ years of experience building scalable data platforms, high-throughput APIs, and cloud-native applications. Currently building data infrastructure at Chicago Education Advocacy Cooperative, transforming education data into research-ready datasets using Azure, Kafka, Databricks, and Spark. Working across large-scale data pipelines, distributed databases, and AI-powered research tools, with measurable impact including 85% fewer duplicate student profiles, sub-second search across 2M+ documents, and dashboard response times reduced to under 200ms.
Previously, at I-Tech Systems, I developed a real-time bidding platform handling 50K+ ad-auction requests per second while maintaining strict latency targets and improving billing accuracy. At Skyline Infrastructure, I built distributed developer tools and Kubernetes-based CI/CD infrastructure, reducing CI execution time from 45 minutes to 12 minutes and average build time by 60%.
My technical expertise spans Python, Java, Go, SQL, Kafka, Spark, Databricks, Flink, PostgreSQL, Redis, Elasticsearch, Kubernetes, Docker, Terraform, Azure, AWS, and distributed systems design. I also work with Azure OpenAI, retrieval-augmented generation (RAG), and MLflow to build practical, data-driven solutions.
Always open to connecting with engineers and technology professionals building impactful products, scalable platforms, and intelligent data systems. 🤝
- Built and deployed a five-node LangGraph agent (extract → triage → retrieve → evaluate → draft, with a real conditional short-circuit) that reads clinical notes, retrieves the governing payer policy via RAG, checks the patient against medical-necessity criteria, and drafts a justified authorization request with citations and a confidence score.
- Modeled all patient data as FHIR resources from day one, the actual industry standard the CMS prior authorization rule is built around, rather than a generic schema, a deliberate choice for domain authenticity.
- Extended with an appeals agent, human-in-the-loop reviewer override, batch processing, a CMS-compliant FHIR Prior Authorization Support API layer, and a PDF-to-policy ingestion pipeline; proved end to end by feeding it a real MRI-brain policy PDF and adjudicating a live claim against it with zero code changes. 🔗 https://github.com/Sakshi3027/prior-auth-assistant | 🌐 prior-auth-assistant-six.vercel.app
- Built the payer-side counterpart to Prior Authorization Assistant, a claims adjudication engine where business analysts define rules directly in Gherkin, a Spring Boot engine evaluates every claim against them, and auditors review decisions through an Angular/RxJS dashboard. Together, the two projects cover the entire healthcare payment lifecycle, before treatment and after.
- Designed a MongoDB document schema for variable-shape claims data and an append-only audit trail logging every adjudication decision and manual override, giving auditors a permanent, provable trail for regulatory traceability.
- Containerized the full stack with Docker Compose for one-command deployment; live on MongoDB Atlas with the API and frontend running as separate Render Docker services. 🔗 https://github.com/Sakshi3027/claims-adjudication-platform | 🌐 claims-adjudication-platform-1.onrender.com
- AI agents work great in the demo, then quietly degrade in production with no one watching. Built a framework that wraps any LangGraph agent, traces every step, and scores output quality automatically with an LLM-as-judge, no human labeling required.
- Implemented drift detection that compares recent scores against a historical baseline and fires an alert the moment quality slips below a set threshold; in its own demo run, it caught a real quality drop (0.57 average against a 0.70 threshold) automatically, before manual review would have caught it.
- Built entirely on free-tier infrastructure (Groq, DuckDuckGo search, SQLite, Streamlit) for $0 total infrastructure cost, purpose-built around the exact failure mode a Forward Deployed Engineer gets called in to solve. 🔗 https://github.com/Sakshi3027/agenttrace | 🌐 huggingface.co/spaces/Sakshi3027/agenttrace
- Built an autonomous agent that wakes up every 6 hours, ingests 500+ signals from 17 sources (RSS, GitHub trending, arXiv) into DuckDB, and reasons through a 4-node LangGraph pipeline, Supervisor → Researcher → Analyst → Judge, with conditional retry when a report scores below 0.7, fully hands-off intelligence generation.
- Diagnosed and fixed a real production hallucination bug: the retrieval layer was passing low-similarity chunks through as trustworthy context, so the model reasoned confidently from noise. Fixed the root cause with concrete similarity thresholds and structured output validators.
- Instrumented full LangSmith trace observability across every node afterward, so any future failure surfaces immediately instead of being discovered by accident; semantic search powered by 384-dim sentence-transformer embeddings in Qdrant Cloud. 🔗 https://github.com/Sakshi3027/signal | 🌐 signal-dashboard-chi-six.vercel.app
- Research Assistant – I worked with a faculty member to make ML experiments run faster. I cleaned up the data pipelines and automated repetitive steps, which cut down experimentation time by about 30%.
- Teaching Assistant – I ran the lab sessions for an ML course around 60 students. I helped them work through the hands-on parts and graded their assignments.
- Data Analyst Intern – I worked with healthcare data, Medicare and Medicaid claims. I wrote SQL queries to clean and validate the data, then built models and dashboards in Python to help the business make better decisions.
- Data Science Intern – I built a model that predicted which customers were most likely to buy, and segmented them so the marketing team could target them better. That ended up increasing conversion rates by 20%.
"The goal is to turn data into information, and information into insight." — Carly Fiorina