Rajath John Bosco
New York, New York, United States
2K followers
500+ connections
About
I build production LLM, retrieval, and analytics systems at JPMorgan Chase. My work…
Experience
Education
-
University of Illinois at Urbana-Champaign
-
-
Specialized in Applied Machine Learning, Data Science, Software Systems, and Analytics.
Relevant Coursework:
• Data Structures & Algorithms
• Programming Analytics & Data Processing
• Applied Machine Learning: Team Project
• Methods for Data Science
• Data Visualization
• Information Modeling
• Data & Statistical Models
• Global Health Informatics
• Sociotechnical Information Systems
• Usability Engineering -
-
-
Licenses & Certifications
-
The Data Scientist’s Toolbox
Coursera Course Certificates
Projects
-
AgentSLA
- Present
Reliability runtime for tool-calling agents with policy gates, numeric verification, execution budgets, append-only traces, deterministic replay for deterministic adapters, and failure analysis.
-
DraftForge
- Present
EAGLE-3 draft-head training and evaluation pipeline. Three A100 seeds measured 68.7% +/- 1.0% held-out greedy agreement. Serving inter-token latency remains unmeasured pending a weight-schema adapter, so no serving speedup is claimed.
-
GoodputLab
- Present
SLO-aware vLLM control plane and benchmark rig. A 54-cell dedicated-H100 sweep reconciled every cell and found that disaggregation slightly improved mean inter-token latency but lost on time to first token and hardware cost for the measured 7B setup.
-
Adversarial Prompt Suite
- Present
LLM red-teaming framework spanning 6 attack categories, with a two-layer classifier (regex heuristics, then an LLM judge for borderline cases) that produces structured coverage reports.
-
Jax Transformer Implementation
- Present
Transformer in pure functional JAX with MHA, MQA, and GQA attention and JIT/vmap/pmap parallelism. Verified MHA/GQA numerical equivalence.
-
Triton Inference Kernels
- Present
Fused softmax and Flash Attention kernels in Triton. 1.5-1.93x softmax speedup and 50x lower VRAM than naive attention at sequence length 2048.
-
LLM Inference Benchmark
- Present
Production-grade benchmarking framework comparing vLLM, TensorRT-LLM, and TGI under realistic bursty traffic patterns.
Most benchmarks use synthetic uniform traffic. Real production is bursty with variable request lengths and queuing dynamics that don't appear in synthetic tests. This framework measures what actually impacts latency and throughput.
Features:
• HDR histograms for accurate p50/p95/p99 latencies
• Realistic traffic profiles (uniform, bursty…Production-grade benchmarking framework comparing vLLM, TensorRT-LLM, and TGI under realistic bursty traffic patterns.
Most benchmarks use synthetic uniform traffic. Real production is bursty with variable request lengths and queuing dynamics that don't appear in synthetic tests. This framework measures what actually impacts latency and throughput.
Features:
• HDR histograms for accurate p50/p95/p99 latencies
• Realistic traffic profiles (uniform, bursty, variable-length)
• Cross-backend comparison on identical workloads
• 245-312 req/s depending on backend configuration
Link - github.com/jrajath94/llm-inference-benchmark -
Tropical Storms
Applied spatial trigonometry, line-intersection algorithms to find possible loops in the storms and usedPyGeodesy to calculate speed, distance, etc. and categorized Atlantic and Pacific hurricanes from US NHC Data.
-
Credit Card Fraud Detection
The project aimed to detect Credit card frauds using a highly imbalanced Dataset with the help of Sampling techniques and regression algorithms .
-
Congestion Control Algorithm using Load Shedding
Basic algorithm that relieves congestion by using primary principles of load shedding.
The algorithm relieves congestion in a network by considering the priority of the packets or the time of generation of the packets based on the type of information being transmitted across the network.Other creators -
Airline Reservation Management System
The project aimed to implement an airline reservation system using MySQL along with a Supporting Front End Web Design.
The project had an admin view and a customer view.
The customer could see all the available flights, reserve tickets (Business/Economy) and pay for the tickets.
The admin could schedule flights ,see reservations , assign pilots, obtain traveler details etc.
Platform: Linux Ubuntu
Languages Used: MySQL (Database Management), PHP, HTML
Tools: MySQL…The project aimed to implement an airline reservation system using MySQL along with a Supporting Front End Web Design.
The project had an admin view and a customer view.
The customer could see all the available flights, reserve tickets (Business/Economy) and pay for the tickets.
The admin could schedule flights ,see reservations , assign pilots, obtain traveler details etc.
Platform: Linux Ubuntu
Languages Used: MySQL (Database Management), PHP, HTML
Tools: MySQL, Xampp.
Other creators -
Vehicle Registration Certificate (RC Book) Management System.
The project was aimed to build a system to manage Vehicale Registration Certificated in India using Object Serialization in Java.
The System had the following features : -
• Add a new Two/Four Wheeler Vehicle Details
• Delete a Vehicle from the List
• Update/ Transfer Details of a Vehicle
• Display the Details of the Vehicle -
American Community Survey
-
See projectPredicted the median household income of American families using supervised Learning models in Python, visualized important factors using Plotly and deployed it using Heroku and Voila.
-
Prediction of Solar Radiation
-
Developed multiple supervised learning models to predict solar radiations based on enriched meteorological data in R.
Other creatorsSee project -
Olympics Analysis(1896 - 2016)
-
Visualized Olympics dataset and explored how home-advantage, physiological, historical, demographical features, affect the performance of a country in the Olympics and predicted the medal tally of a country using HDI, Population, GDP, GDP per-capita using Random Forests with an accuracy of 96%.
Other creatorsSee project -
Bike Review Analysis in Social Media
-
The project involved obtaining the bike specifications from a relevant source(Website).
Later we extracted reviews from Facebook comments and tweets from various users regarding a particular bike model.
Python NLTK was applied on the reviews to obtain the sentiment of the review and the overall user reception.
The model can be extended to other products as well.Other creators
Languages
-
Kannada
Native or bilingual proficiency
-
English
Native or bilingual proficiency
-
Hindi
Full professional proficiency
Organizations
-
Linux Campus Club
Executive Committee Member
-
Other similar profiles
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content