Working at the intersection of AI and Backend Engineering

Tech Stack : Python, FastAPI, PostgreSQL, GCP, Redis, Prompt Optimisation, Embedding models, Huggingface, PyTorch

1. Developed Automated Construction Estimates services which was responsible for giving accurate vendor item recommendations leading increased product adoption and usage.

2. Developed various algorithmic techniques for automated symbol detection in legend tables in construction takeoffs.

3. Worked on developing the Data Pipeline for efficient rendering of engineering diagrams into a machine readable format

4. Implemented Supervised Finetuning of Llama 3.1, 3.2 models for duct detection.

5. Deployed AI based services and solutions on FastAPI servers in production.

Web Entity Extraction and Validation at large scale using Deep Learning and LLMs

• Designed and developed H1’s Entity Validator service based on FastAPI, Flask and

Transformers framework for efficient data validation and successfully deployed in Web Extraction Pipeline on AWS to improve

data validation efficiency at large scale, resulting in reduction of turn-around-time of 95%.

• Designed and developed the PySpark Web Extraction pipeline for entity extraction having multiple services which replaced the legacy extraction framework leading to increase in throughput of 600%.

• Engineered algorithms for reverse mapping, voting to identify common XPaths for increased efficiency in extraction by 30%.

• Development of Cleanup module which resulted in increased quality of production data and reduced data issues by 40%.

• Worked on Prompt engineering and fine tuning for OpenAI GPT-3.5-turbo based Entity Extraction increasing recall by 16% and precision by 11%.

Skills: RDBMS · MySQL · Back-End Web Development · Continuous Integration and Continuous Delivery (CI/CD) · PySpark · SQL · Data Structures · BERT · Apache Spark · Machine Learning Algorithms · Data Engineering · Mathematics · Pandas · NumPy · Data Science · Computer Science · Natural Language Processing (NLP)

• Optimized processor resource consumption of Samsung’s Bluetooth algorithm used in flagship models: Scalable Speech Codec

(SSC) leading to 15% MIPS (Millions of instructions per second) reduction.

• Upgraded Samsung’s Diamond-Voice Module from Q15 to Q24, Q31 format and optimized it, leading to 5% MIPS reduction.

• Optimized SSC Library Audio Codec memory usage for Hifi3 Low Memory chip leading to 40% memory usage reduction.

• Optimized processor resource consumption of SSC Encoder and Decoder Legacy code resulting in 9.1% reduction in MIPS.

• Authored Conference Paper titled “Combining Pixel Selection with Covariance Similarity Approach in Hyperspectral Face Recognition” accepted in the 44th Annual Conference of the IEEE Industrial Electronics Society

• Authored Conference Paper titled “Fast Linear Unmixing of Hyperspectral Image by Slow Feature Analysis and Simplex Volume Ratio Approach” accepted in the 39th Annual IEEE Geoscience and Remote Sensing Society symposium

• Authored Journal Paper titled “Library Aided Bilinear Unmixing of Hyperspectral Image using Subspace Clustering and Multistep Pruning” accepted in Journal of Applied Remote Sensing