Sanket Shigaonkar

Software Engineer · ML & AI Systems

Building reliable ML systems, RAG applications, and backend tools.

Open to Software Engineering, Machine Learning, and Applied AI roles.

New York, NY · Open to relocate

My Resume

About Me

I’m a software engineer working where machine learning, backend systems, and product engineering meet. I recently completed my MS in Computer Science at the University at Buffalo, where I built and delivered a citation-bearing RAG backend spanning seven years of academic content.

Today, I work on smartphone-based anemia screening at Iron Strong Health Initiative. My focus is making model results trustworthy: finding data leakage, building reproducible evaluation systems, testing domain shift, and measuring how models behave under realistic input variation.

Outside work, I build practical AI and developer tools—from an LLM job-matching workflow to a publicly released Firefox extension. Away from the keyboard, I’m usually playing basketball, watching the Warriors, or listening to The Weeknd.

Some of the technologies I have been working on recently:

  • Python
  • PyTorch
  • FastAPI
  • LangGraph
  • PostgreSQL
  • Qdrant
  • pgvector
  • Docker

Experience

  1. Software Engineer, Machine LearningIron Strong Health Initiative, Inc.

    Jul 2026 — Present

    South Orange, NJ

    • ›Rebuilt evaluation for a smartphone anemia classifier after uncovering subject-level leakage, establishing subject-grouped cross-validation across 216 subjects and a reproducible three-seed protocol.
    • ›Evaluated about 25 model and training configurations; developed an interpretable color baseline reaching 0.889 pooled AUROC and a color/deep ensemble reaching 0.903 pooled AUROC.
    • ›Diagnosed country-associated domain shift and built crop-robustness tests to distinguish ranking degradation from threshold drift.
    • Python
    • PyTorch
    • scikit-learn
    • Medical Imaging
    • ML Evaluation

  2. AI Engineer / Research AssistantUniversity at Buffalo

    Feb 2026 — Jun 2026

    Buffalo, NY

    • ›Built and delivered a citation-bearing RAG backend with LangGraph, Qdrant, FastAPI, and Docker Compose over 1,300 web pages and 400 handbook pages spanning seven years.
    • ›Added year-filtered retrieval, bounded query rewriting, and explicit fallback, improving year-specific correctness from 0/13 to 12/13 and reducing unsupported answers from 9/49 to 1/49 on a coordinator-built evaluation set.
    • Python
    • LangGraph
    • Qdrant
    • FastAPI
    • Docker
    • RAG

  3. ML / Application Engineering InternDentite

    Jan 2025 — May 2025

    Buffalo, NY

    • ›Created a PHI-free synthetic evaluation corpus of 1,000+ insurance-card images across six payer templates for repeatable multimodal-LLM benchmarking.
    • ›Compared GPT-4, Gemini, and Claude Sonnet on accuracy, cost, and latency, selecting Claude at 93% aggregate field-level exact match across five fields on the synthetic ground-truth set.
    • ›Built a React Native capture client and async FastAPI extraction service, then integrated the structured output into the existing claims workflow.
    • Python
    • FastAPI
    • React Native
    • TypeScript
    • LLM Evaluation

Projects

  • ApplyTrak — LLM Job-Matching WorkflowLive Demo

    • ›Built a publicly deployed workflow that extracts structured requirements from job posts, deduplicates reposts with pgvector, and reranks candidates against a resume.
    • ›Implemented provider-agnostic LLM contracts, embedding provenance, content-hash caching, rate limiting, and idempotent resumable processing.
    • FastAPI
    • PostgreSQL
    • pgvector
    • Pydantic
    • Vercel
    • Neon

  • Claude Chat Exporter — Firefox Extension

    • ›Built and published a privacy-oriented Firefox extension that exports named and incognito Claude conversations to local Markdown without a developer backend or telemetry.
    • ›Fixed long-chat truncation caused by UI virtualization through incremental top-to-bottom harvesting and ordered merging of partially mounted views.
    • JavaScript
    • Firefox WebExtensions
    • DOM APIs
    • Turndown

  • Accio AI — QLoRA-Tuned RAG Assistant

    • ›Fine-tuned Qwen3-4B-Instruct with 4-bit QLoRA and Flash Attention 2 on 25,253 formatted instruction samples in a completed packed one-epoch SFT run.
    • ›Built a local PDF-grounded generation pipeline with PyMuPDF, MiniLM embeddings, and exact FAISS retrieval to inject the top three source chunks.
    • Qwen3-4B
    • Transformers
    • PEFT
    • TRL
    • FAISS
    • Streamlit

  • DeepDish — Food Detection & Segmentation Pipeline

    • ›Trained and evaluated YOLOv8s and Faster R-CNN food-and-coin detectors; YOLOv8 reached 0.9675 mAP50 and 0.8822 mAP50–95 across 1,733 test images.
    • ›Implemented OpenCV GrabCut segmentation over detected food crops for the team’s downstream volume-estimation pipeline.
    • Python
    • PyTorch
    • YOLOv8
    • Faster R-CNN
    • OpenCV
    • COCOeval

  • Smart YOLO Intersection Traffic SystemPublished Patent Application

    Named co-inventor on Indian Patent Application No. 202321005578 for an intersection traffic-management system combining YOLO-based vehicle detection, traffic-density forecasting, and adaptive signal control. Published February 17, 2023.

    • Computer Vision
    • YOLO
    • Residual LSTM
    • Adaptive Signal Control