Niare Doyom
Junior AI Engineer at Morgan Stanley Capital International
Passionate about AI, machine learning, and building scalable data extraction platforms. Specialized in transformer architectures, NLP, and cloud-based solutions.
Education
B.Tech in Engineering Physics
Nov 2020 - May 2024Indian Institute of Technology Bombay
Relevant Courses:
- Deep Learning for NLP | Prof. Pushpak Bhattacharyya | Achieved perfect 10/10 grade
- Microprocessors | Prof. Pradeep Sarin | Achieved perfect 10/10 grade
Work Experience
Junior AI Engineer
Morgan Stanley Capital International June 2024 - Present- Worked with a team of 16 to build and maintain the in house Data Extraction Platform for Financial Documents
- Designed and deployed in a team of 3 a streaming extraction pipeline using GCP Dataflow and Google Pub/Sub, enabling seamless high-throughput processing 3,600+ documents daily in production
- Developed, in a 3-person team, a Data Extraction API (FastAPI + Spanner DB) supporting update, view, and addition of data for HITL verification; achieved strong consistency and latency under 500 ms for 1,000+ records and handled 250k+ requests daily in production
- Independently designed dataset-specific prompts for experimenting with multiple LLMs (Google, Anthropic, OpenAI), achieving 73%+ precision and 86%+ recall across 26 datapoints
- Lead a team of 4 developing prompts for several datasets enabling extraction for 60+ datapoints with 72%+ precision & 80%+ recall
- Contributed to a prompt generation agent that generates extraction prompts from datapoint names and financial methodology documents, reducing manual effort and achieving 40%+ precision and 60%+ recall
- Worked closely with a VP to build the Full Extraction feature, enabling users to manually trigger extraction across multiple pages and map outputs to relevant datapoints, eliminating the need for dataset-specific prompts
Data Analyst
Yun Solutions LLP Aug 2023 - Nov 2023- Brainstormed in a team of 5 to develop a multimodal predictive model which uses alternative data for stock prediction
- Curated a quality dataset with keyword filters, sentence similarity, and GPT-3.5 for better data enrichment and validation
- Performed sentiment analysis with Roberta and FinBERT, finding 85% correlation between tweet sentiment & stock prices
- Analyzed and visualized trends with Matplotlib, Plotly, and Power BI to present clear insights and patterns
- Implemented agglomerative clustering of review replies to identify automated response to app reviews
Technical Skills
Languages
Frameworks & Tools
Software
Hobbies
Key Projects
Unlimiformers
Dec 2023 - Apr 2024- Applied a approach for extending the context length of encoder-decoder transformers by implementing "Unlimiformers" paper
- Utilized kNN index for cross-attention computation, enabling the processing of practically unlimited input sequences
- Perform chunking for encoding the input sequence to process sequences for which token length exceed the context-size
- Demonstrated the effectiveness of Unlimiformer on various long-document and book-summarization benchmarks, handling inputs as long as 500k tokens from the BookSum dataset
Shakespearean Text Generation
Aug 2023 - Nov 2023- Developed a custom transformer architecture referring Attention is All you need Paper from the ground up using PyTorch
- Implemented multi-headed self attention mechanisms entirely from scratch to enhance model capabilities
- Specialized in training the transformer model on the Tiny Shakespeare dataset
Autograd Engine
July 2023 - Aug 2023- Created a compact Autograd engine designed to facilitate backpropagation through dynamically built Directed Acyclic Graphs
- Implemented reverse-mode auto differentiation, fostering efficient gradient computation over dynamically built DAGs
- Employed a strategic approach by fragmenting each neuron into elemental operations such as tiny adds and multiplies
- Incorporated 20+ different operations for nodes improving the computational efficiency of gradient determination