I build the unglamorous half of AI products — the FastAPI service, the SQL pipeline, the retrieval that decides which document the model is allowed to answer from. Two and a half years of it, mostly for a banking client.
Most of an AI feature is not the model. It is the request path around it — and that is the part I own.
RequestFastAPI route, validated JSON
RetrieveSQL + vector search
Groundonly the sources that matter
Answertraceable back to text
The same path runs under ServiceIQ at UBS: service data in, grounded analysis out, sources attached.
Experience
Junior AI Engineer
May 2025 — Present
UBS Investment Bank client engagement · via Atyeti IT Services · Pune
ServiceIQ — built Python APIs and SQL-backed data-processing workflows that feed enterprise service data into dashboards, as part of a larger team-built platform.
Contributed document-grounded LLM insights to ServiceIQ using retrieval, so generated analysis is anchored in relevant source material rather than the model's memory.
Built backend automation and pipelines for PDF, Excel and enterprise datasets, cutting manual processing time.
Shipped production-ready services with client stakeholders: Dockerised deployments, API validation, RCA and debugging to keep them stable.
Data Science Intern
Jan 2024 — Feb 2025
Revealit Solutions · Bengaluru
Developed REST API endpoints in Python with JSON validation, request parsing and structured error handling.
Built database-interaction modules and automated data-processing workflows with hash-key-based record management.
Improved backend data integrity, reducing processing inconsistencies.
Projects
AI job recommendation engine
FastAPI backend for an AI-driven job matching platform, with dynamic SQL query generation behind an agentic chatbot.
Architected the backend: request models, query builders and response shaping for a chatbot that has to answer in structured data, not prose.
Dynamic SQL generation so the chatbot's questions become queries instead of hard-coded filters.
Python · FastAPI · SQL · LLMs
PDF redaction module
A pipeline that finds and removes sensitive information in scanned and dynamic PDFs — and actually deletes it, rather than drawing a box over it.
OCR to read scanned documents, then detection of sensitive fields before flattening the PDF so removed content cannot be recovered.
Resolved Linux compatibility issues in the toolchain — the part that makes a demo into something a team can run.