Rishab Pal

Founded an AI engineering studio building agentic, vision, and generative AI systems for startups and enterprises. Led projects including a text-to-SQL agent, LLM fact-checking, MCP migration, CCTV vision, and generative media.

AI engineering studio that builds production agents, retrieval and vision systems for startups and enterprises. I own each engagement end to end: discovery, architecture, build, deployment and support.

• Text-to-SQL agents: Built a LangGraph agent that answers business questions over a Databricks warehouse by running each question as a parallel DAG of subqueries. Verified SQL templates cover know

Led data science for search, recommendations, and pricing. Improved search CTR by 50% with semantic matching and learning to rank, won back 20% of churned customers, and generated GenAI campaign creative.

Data science across eCommerce search, pricing, customer lifecycle and marketing creative.

• Search rebuild: Re-architected the eCommerce search stack for organic and ad merchants. Custom deep semantic matching models and learning to rank lifted click-through 50%, and the rebuild added 10% incremental revenue.

• Keywords and scale: Built a new search keyword generation approach and moved retrieva

Improved roof line detection accuracy from 28% to 59% using a Transformer on aerial imagery. Implemented segmentation with 86% mIoU and reduced latency by 25% with INT8 quantization on TensorRT and Triton.

Computer vision for automated roof measurement of US homes from aerial imagery.

• Line detection: Built a Transformer-based model that detects roof, wall and ground lines in aerial images. Better data quality, augmentation and fine-tuning doubled accuracy, from 28% to 59%.

• Facet segmentation: Trained segmentation models for roof and wall facets that hold up when parts of the roof are obstructe

Built a Hinglish assistant for 400M users with a Transformer chatbot at ~85% accuracy. Rebuilt face recognition and deduplication for KYC fraud defense, cutting manual review by 25%.

Product Engineering team.

• Conversational assistant: Built a Transformer-based chatbot that resolves customer queries in English and
code-mixed Hinglish for 400M+ monthly active users. Deployed at commercial scale, it handles thousands of conversations a day at about 85% accuracy.

• Face dedupe KYC: Led the rebuild of the face deduplication platform used in KYC, upgrading the face recognition,

Worked on FortressIQ's process intelligence platform (later acquired by Automation Anywhere), which applies NLP, computer vision, imitation learning and few-shot learning to digital transformation.

• Privacy Enhanced Gateway: Completed and deployed the platform's Privacy Enhanced Gateway on a custom OCR stack: region proposal and text recognition models plus BERT-based semantic correction.

• OCR

Developed custom text detection and BERT-corrected recognition that beat Google Vision OCR by 12% on ICDAR 2013. Built document AI for hiring with LayoutLM, NER, and embedding search.

Emerging Technology and Innovation Centre. Led and owned AI Conductor, the onboard AI for Yamaha's self-driving Public Personal Mobility vehicle, and presented it in a live demo at CES 2019 in Las Vegas.

• Gesture control: Designed a ConvLSTM hand gesture recognition model, tuned with Bayesian optimization.

• Face recognition: Built single-shot face recognition with SSD MobileNet for detection a

Built and owned the onboard AI Conductor for Yamaha's self-driving vehicle, including ConvLSTM gesture control, face recognition, and Kalman filter collision warning. Led retail sales forecasting with ARIMA and Holt-Winters models.

• Sales forecasting: Led a team that proposed and built retail sales forecasting with ARIMA, exponential smoothing and Holt-Winters models, lifting quarterly sales 5.35%.

• Drone video: Built a YOLOv2 object detection model to count cars in drone footage.

• Generative design: Used a Wasserstein GAN to synthesize new vehicle designs from sample designs.

• IoT: Worked on Arduino and Raspberry Pi

Built models to predict energy consumption in a low-energy house.

• Modeling: Combined PCA, support vector machines and ensemble methods (Random Forest, XGBoost) to model consumption patterns.

• Tuning: Used Bayesian optimization to tune the hyperparameters of the neural network models.

Sep 2019

Apr 2018

Grade: 8.74/10

Pursued a Bachelor of Technology in Computer Science and Engineering, achieving a CGPA of 8.74/10.

Feb 2018

Feb 2018

Dec 2018

Sep 2018

May 2018

Mar 2018

Apr 2018

Mar 2018

Feb 2018

May 2017

May 2016

May 2016

Apr 2015

Jun 2015