Software Engineer (AI) at TSS Consultancy, Rajkot. I build production computer vision systems and multi-agent LLM pipelines.
Most of my day job ships behind a RegTech product and stays private, so this profile is mostly the work I can show: agentic pipelines with real evaluation harnesses, and CV work I rebuild in the open.
Current focus: multimodal vision-agent systems, where video understanding meets agentic reasoning.
KYC and compliance infrastructure for a WestBridge-backed RegTech product (SOC 2 Type 2, ISO 27001).
- Face liveness detection API running in production, tuned for sub-second p99 under real request load
- Object detection and multilingual OCR across 8 Indian languages
- Ongoing research into deepfake and forged document detection
SmartIngest — agentic document intelligence
Upload an invoice, contract, resume or ID. It classifies the document, extracts structured fields, validates them against business rules, and routes it to auto-approve, review, or reject.
LangGraph orchestration, Gemini vision, Pydantic-typed extraction, prompt-injection and PII guardrails, grounding checks with a confidence-triggered retry loop, deterministic YAML routing, LangSmith tracing end to end.
Python 3.12 LangGraph FastAPI Gemini Pydantic Docker LangSmith
72 tests, eval harness wired as a CI gate. 90-second demo
RecruitRadar — multi-agent recruitment screener
Drop in a job description and a folder of resumes. Get a ranked shortlist with auditable 0 to 100 scores, tailored interview kits, and drafted outreach.
Five CrewAI agents with an embedding pre-filter. Scoring is computed in Python with weighted sums, never emitted by the model, and every matched skill has to cite a verbatim resume quote verified with rapidfuzz.
Python CrewAI Gemini LiteLLM rapidfuzz SQLModel SQLite
Full walkthrough
More at kushagrapandya.in.
Languages Python, Java CV PyTorch, OpenCV, YOLO, Vision Transformers, ONNX LLM / agents LangGraph, LangChain, CrewAI, LangSmith, ChromaDB Backend FastAPI, Flask, Docker, Azure, SQLModel
- 2 IEEE conference papers, including Barcode and QR Code Object Detection: An Experimental Study on YOLOv8 Models (INNOCOMP 2024)
- 4 patent applications filed with IP India, 1 granted: a navigational aid device for the visually impaired
- Earlier research on sign language translation and myoelectric prosthetic control
- ORCID: 0009-0007-7695-335X
I publish roughly weekly on Medium, mostly notes from building agent systems: structured output, middleware, evaluation, and what breaks in practice.
B.Tech, Computer Engineering (AI specialization), Marwadi University, 2024. CGPA 8.51.
Portfolio · LinkedIn · Medium · Hugging Face