Projects - Sahasra Ranjan

Founding AI Engineer (2nd engineer) | Sunnyvale, CA | Aug 2025 - Present

Architected Claude Agent SDK integration: designed unified subagent system with task delegation, native streaming, and 70/30 message compaction. Core infrastructure powering all AI features.
Built Odoo ERP connector from scratch: schema introspection, field customization agents with infinite loop prevention, CSV validation, and data migration workflows. Reduced 6-week enterprise migrations to days.
Designed streaming citation system for Knowledge Q&A: O(n) grounding parser, real-time inline references, and artifact resolution. Built full-stack end-to-end (database migrations, backend services, frontend UI).
Implemented AI-enhanced PDF processing: Claude vision + Gemini for image descriptions, two-pass extraction (instant feedback + async deep analysis), and original-quality image extraction from complex documents.

Co-founder, AI Math Tutor for K-8 | San Francisco, CA | 2024 - 2025

Co-founded AI tutoring startup to democratize math education; built platform where kids create adventures and characters with math woven into storylines. Piloted with 150+ students achieving 95%+ retention over 6 months.
Managed team of 4 engineers and 6 tutors; conducted 50+ hands-on sessions to understand what kids and parents care about. Discovered younger students need emotional connection before cognitive engagement.
Human-in-loop model proved unscalable; led to insight that some gaps between models and humans are structural, not accidental, motivating deeper interest in AI alignment and robustness.

Guide: Prof. Preethi Jyothi, CSALT Lab | B.Tech. Project at IIT Bombay in collaboration with Google Research India | 2022 -

Currently working on a generalised framework for code-switched text generation, with multiple possible constraints on formality, politeness, semantic similarity, etc – test tasks on English-Hindi and English-Spanish
Exploring multiple MLM techniques based on word language to improve code-switched text generataion model.
Working on Frequency MLM technique which improves the text generation without the word language information

Software Engineering Internship | Rubrik, Bangalore, India | 2022

Worked on a GC algorithm for a snapshot tool in a cloud data backup and recovery system using Go Lang & Korg.
Developed a mechanism to track expired critical data snapshots which were blocking 30-40% of the database table.
Ideated & implemented a framework to periodically find unnecessary snapshots and remove them from the database

Prof. Martin J. Lercher, Alexander Kroll | Research Internship at HHU, Dusseldorf, Germany | 2021

Worked on a BERT based model on PyTorch to create high dimensional representations for enzyme sequences for the prediction of enzyme-substrate binding (binary classification task)
Adapted ESM model to build an end-to-end model which outperformed the existing SOTA models for the task
Improved accuracy for the task was achieved from 79.8% to 87.5% with the new representations that were created with parallel training on 8 GPUs for 10 epochs

Guide: Prof. Ajit Rajwade | Course Project: Advanced Image Processing

Applied Adaptive Median Filtering algorithm for removing impulsive noise from the spatial domain of the video
Implemented Three Step Cross Search on MATLAB for organising similar patches across the temporal domain
Adapted Fixed Point Iterative Algorithm for reducing Poisson and Gaussian noise from the image sequence

Guide: Prof. Vinay J. Ribeiro | Course Project: Introduction to Blockchains, Cryptocurrencies, and Smart Contract

Build a discrete-event simulator for P2P cryptocurrency network on python to study the blockchain network
Implemented selfish mining attack proposed by Eyal and Sirer in the paper ”Majority is not Enough”
Build a layer-2 Decentralised APPlication (DAPP) on top of Etherium blockchain using Solidity and Ganache

Guide: Prof. Biswa | Course Project: Computer Architecture

Worked on P-OPT algorithm proposed in the paper "Practical Optimal Cache Replacement for Graphs"
Used Champsim cache simulator to extract data and implement the algorithm for a fixed graph benchmark
Achived an improvement in LLC load hit rate by a factor of 3x as compared to LRU replacement policy

Institute Technical Summer Project | Institute Technical Council

Implemented a CNN and LSTM based model on PyTorch for generating LaTeX expression of the input equation
Adapted the OpenAI problem statement and used im2latex-100k dataset to achieve a BLUE-4 score of 38.82
Deployed the model on a Django based interactive web application and integrated the Django-TeX parser

Guide: Prof. Amitabha Sanyal | Course Project: Software and System Labs

Created a web platform providing an in-browser IDE and support for user organised coding competitions
Used Django, Angular and PostgreSQL to create an interactive user interface with secure user-authentication
Implemented user directory separation, multiple language support and isolation using Docker for security.

Guide: Prof. Ajit Rajwade | Course Project: Advanced Image Processing

Adapted publication from ICCV’11 to reconstruct spatial and temporal domain of the video from coded snapshot
Implemented Orthogonal Matching Pursuit algorithm for sparse reconstruction to achieve RMSE of 0.03301

Guide: Prof. Ajit A. Diwan | Course Project: Data Structures and Algorithm

Created a Region quad-tree class in C++ to store grey-scale images with highly optimised space complexity
Implemented optimised algorithms to allow for intersection, overlap, resize, complement, and extraction of images

Guide: Prof. Vinay Ribeiro | Course Project: Computer Networks

Implemented client and server using Socket Programming in C, to send files using different variants of TCP
Used Bash to automate experiments and generate plots for comparing throughput, delay, and packet loss
Recorded network traffic using Wireshark and analysed window scaling graphs for TCP Cubic and TCP Reno

Guide: Prof. Ashutosh Gupta | Course Project: Logic for Computer Science

Encoded moves of the mastermind game into an SAT problem and solved using conflict driven clause learning
Implemented a solver in Python using z3py library which was robust to the other player lying up to 30% of the time

Database and Backend Internship

Lead the backend development in a telemedicine startup to schedule non-contact consultations for patients
Implemented a database for storing encrypted information of doctors and patients on AWS using PostgreSQL
Worked on a Django based backend enabling end-to-end features including in-app audio and video calling