Shreya Kommuri
Senior AI Engineer at Propheus
About
I'm an AI/ML Engineer with 2+ years building production-grade GenAI and computer vision systems. At Propheus, I build computer vision pipelines that extract structured visual features from street-level video imagery, fine-tuning vision-language models and detection architectures to support real-world scene understanding at scale. Previously at Zikhara AI, I developed GenAI-powered design automation systems using multimodal LLMs, reinforcement learning, and agent orchestration, with a track record of deploying scalable solutions that reduced manual operations by 70%+.
I'm comfortable in ambiguous spaces where the solution isn't obvious. I shape fuzzy problems into concrete architecture and evaluating whether we need LLMs or heuristics, optimizing for latency and cost, and building monitoring to track quality over time.
I'm interested in AI alignment and interpretability research. While my current focus is engineering, I explore these topics through writing and personal projects, thinking about how we can build AI systems that remain beneficial as they become more capable.
Experience
Senior AI Engineer · Propheus
Built computer vision systems to extract structured visual features (signboard regions, text, object boundaries) from street-level video imagery, enabling automated business/location data extraction at scale. Fine-tuned MiniCPM-V via SFT/DPO for entity name and text recognition, and fine-tuned YOLO (OBB variant) on custom-labeled data with cross-frame object tracking to reduce duplicate detections and per-frame inference cost.
Benchmarked multiple detection and vision-language architectures against precision, recall, and accuracy criteria to guide production model selection, and built data curation pipelines and evaluation frameworks with task-specific metrics.
AI Developer · Zikhara AI
Leading development of GenAI-powered design automation using multimodal LLMs for text, image, and layout understanding. Built computer vision pipelines that transform static product images into editable design templates, reducing manual work by 70% and saving designers 3-5 hours per campaign.
Developed SVG parsers for Canva and Adobe integrations, built a Photoshop plugin for AI-generated design layers, and implemented monitoring workflows with model evaluations and fallback mechanisms. Fine-tuned open-source LLMs including LLaMA and Mistral using LoRA for e-commerce content generation.
Projects
Multi-Agent Task Solver (MATS)
Built AI orchestration system that converts natural language business requests and CSV data into validated insights and visualizations, automating manual data analysis workflows.
Designed 4-agent architecture (planner, analyst, executor, chart generator) with CSV schema validation to prevent LLM hallucination on non-existent columns; implemented correction flows for mismatched requests
Automated Product Listing System
Designed an end-to-end ML pipeline that converts raw product images into Amazon-compliant listings by combining computer vision with LLMs. The system performs category classification, schema extraction, and dynamic prompt generation while parsing 700+ category-specific Excel schemas to enforce compliance rules.
Developed production-grade LLMOps workflows with fallback mechanisms, A/B testing hooks, and data validation to ensure scalability and reliability. Built to handle real-world e-commerce constraints where accuracy and compliance are critical.
QuantEase
Built a full-stack LLM evaluation platform to systematically compare cloud models against locally-quantized variants (GGUF format), enabling users to find optimal speed-quality-size trade-offs for production deployment. Designed and implemented task-aware evaluation pipelines for three domains: text generation, classification, and RAG (Retrieval-Augmented Generation), each with specialized metrics and validation logic.
The backend uses FastAPI with Celery workers for distributed task processing, integrating llama.cpp-python for local GGUF inference and Groq API for baseline comparisons. Implemented comprehensive quality metrics including BERTScore for semantic similarity, task-specific accuracy measurements, and an optional LLM-as-judge system using Llama 3.3 70B for qualitative assessment. Built custom prompt formatters for different model families and quantization-aware evaluation logic that handles edge cases like polysemantic outputs and hallucination detection in RAG tasks.
Created a React + TypeScript dashboard with interactive visualizations showing trade-off curves (model size vs quality, inference speed comparisons), side-by-side output comparisons, and per-sample divergence analysis. The system provides actionable recommendations based on use case requirements, helping teams reduce inference costs by up to 90% while maintaining acceptable quality thresholds. Demonstrates practical MLOps patterns for safe quantization evaluation with PostgreSQL for experiment tracking, Redis for task queuing, and Docker containerization for reproducible deployments.
Writing
Notes: Zoom In circuits Article
Reading old interpretability papers chronologically, pretending I don't know what happened next
Why I'm Spending My Weekend Cleaning CSVs
Phase 1: Data Pipeline – Loading, cleaning, validating, and engineering features from raw ESG data.
Context: I'm building a production-grade ML API to predict company ESG scores using classical ML models. Not that anyone asked for it, just for funsies
Talks
Engineering Custom Diffusion Workflows with ComfyUI: A Developer's Guide
A technical talk on building production-grade diffusion workflows with ComfyUI, covering architecture decisions, optimization techniques, and lessons learned from deploying AI-powered image generation systems at scale.
Contact
Email: shreyakommuri2000@gmail.com
GitHub: github.com/winnie-wildin
LinkedIn: linkedin.com/in/shreyakb20