Teams that scale past calling a provider directly from the app run into unpredictable cost, no visibility into failures, PII/prompt-injection exposure, and no resilience when a provider degrades — usually all at once.
Prism is a self-hosted LLM gateway and control plane that sits between applications and providers — OpenAI, Anthropic, self-hosted Ollama. A one-line base_url swap, no other code changes. It adds cost visibility, automatic failover through a hand-rolled circuit breaker, PII and prompt-injection guardrails, exact-match caching, and full observability, backed by Postgres with pgvector, Redis, and Prometheus/Grafana/Jaeger.
245–259 req/scombined throughput · 0% unintended errors — main load test
0.00%fast-model failure rate during a simulated total Ollama outage — 1 stray timeout in 31,488 requests
FIG. 01ANominal request path — the default straight-through flow. Cache-hit bypass and failure-retry paths are not shown here; see the complete flow for those.
Endoscopic procedures produce a continuous camera feed with no automated way to flag stone material in real time — detection still depends entirely on the surgeon's eye.
Ongoing computer vision research at NYU's FAMS Lab, contributing to a real-time detection pipeline over a continuous endoscopic camera feed inside a simulated renal environment — water and calcium-based model kidney stones standing in for real tissue and stone material during development, as part of the lab's Vertically Integrated Project.
FIG. 08Experimental pipeline
Current focus
Continuous video ingestion
Real-time stone identification
CV pipeline development
Stack
PyTorch / OpenCV / Computer Vision
Case Study / 003Software Systems · 2026
Faculty Operations Platform
500+ NYU faculty ran records management, workflow routing, and approvals through manual, form-based processes with no shared system of record.
A cloud-integrated Django platform replacing manual faculty-affairs paperwork with structured workflows, approval-chain tracking, and an AI-powered FAQ chatbot — compatible with NYU Box, AWS S3, and Azure Blob Storage. Twelve Django apps split by domain, covering the core platform, workflow engine, approvals, notifications, document storage, audit trail, chatbot, workload tracking, data export, tenure and promotion, faculty records, and access management.
Verified figures
12Django apps, split by domain — see module map
306automated tests
29test files
FIG. 09Functional index — 12 Django apps grouped by function for presentation; not a dependency graph.
Airspace congestion has to be understood while it's happening, not after the fact — which means ingesting live flight telemetry and scoring risk in-stream, not in a nightly batch job.
A streaming pipeline that ingests flight telemetry from the OpenSky API (plus historical replay) into three Kafka topics — flight-stream, flight-metrics, flight-aggregates — processes it in Spark Structured Streaming for risk scoring, anomaly detection, and spatial grid aggregation over 30-second tumbling / 10-second sliding windows, and fans results out to InfluxDB for real-time queries, MongoDB for historical lookups, and HDFS for batch archival, visualized in a live Streamlit dashboard.
Verified figures
475,000+ recordsgenerated historical dataset modeling realistic OpenSky-shaped traffic, used for testing at volume — live ingestion pulls the real OpenSky API
Open-source offline RAG chatbot with semantic document search using FAISS vector indexing and LangChain retrieval chains, supporting PDF/DOCX ingestion across 1,000+ documents with sub-2-second latency. Modular pipeline with configurable embedding models, chunking strategies, and reranking.
Take-home assessment for Teiko — analysis pipeline and interactive dashboard for a clinical-trial immune cell population dataset. Converts per-sample cell counts to relative frequencies across 5 immune cell populations, statistically compares treatment responders vs. non-responders with FDR-corrected significance testing, and breaks down baseline cohorts by project, sex, and response. Reproducible SQLite-backed pipeline with a 15-test suite and a live dashboard.
Take-home assessment for Juspay — a concert ticket booking prototype with a real payment integration against Hyperswitch's Unified Checkout sandbox. Server-side amount computation that never trusts client-submitted prices, Redis-backed atomic inventory holds with lazy expiry, and webhook-safe payment-intent cancellation to prevent overselling seats during the 10-minute checkout hold.
CNN-based gesture-to-text translation system improving accessibility for the hearing-impaired. Enhanced spatial feature learning with skeleton-plotted image augmentation. Published in IEEE Xplore 2024.
LLM-driven conversational therapy agent using GPT-4, NLP, and sentiment analysis. Analyzes emotional state during live interactions and implements sentiment scoring + trend tracking.
Automated face/object detection pipeline using AWS Rekognition with intelligent metadata tagging. Enables semantic queries for image search via REST APIs and event-driven inference.
LLM-powered research paper summarizer and Q&A system. Ingests academic PDFs, generates structured summaries, and enables natural language querying over paper content via RAG.
LLMs / RAG / NLP / Python / Streamlit
Smart Door Lock System
Serverless IoT security system built on AWS. Uses Lambda, DynamoDB, API Gateway, and S3 for a fully managed, scalable smart lock with remote access control.
Cloud + blockchain hybrid file storage and verification system. Files stored on cloud with cryptographic proofs anchored on Ethereum for tamper-proof verification.
Blockchain / Ethereum / Cloud / Solidity / AWS
DiningBot on AWS
Serverless dining recommendation chatbot built on AWS. Uses Lex for natural language understanding, Lambda for business logic, and DynamoDB to serve personalized restaurant suggestions.
Contributing to a real-time detection pipeline over a continuous endoscopic camera feed, as part of a Vertically Integrated Project.
Developing and validating detection in a simulated renal environment — water and calcium-based model kidney stones — before any real-tissue application.
Computer VisionPyTorchOpenCV
Software Engineer — Faculty Operations Platform
NYU — Office of Faculty Affairs
Jan 2026 – Aug 2026
New York, NY
Leading system design for a cloud-integrated digital platform compatible with NYU Box, AWS S3, and Azure Blob Storage, serving 500+ faculty with automated records management, workflow orchestration, and an AI-powered FAQ chatbot.
Deploying workflow automation pipelines using cloud-based orchestration (AWS, Azure Logic Apps) to digitize manual form processes with routing, compliance checks, and approval chain tracking.
Designed and deployed a fully offline RAG-based AI chatbot using LangChain, vector embeddings, and semantic retrieval workflows, enabling secure knowledge access across 1,000+ enterprise documents for 50+ field team members.
Built automated document ingestion and semantic indexing pipelines, reducing information retrieval time by ~70% compared to manual search.
Enhanced NLP query interpretation in Power BI Q&A using custom semantic models; integrated Azure AI and Python NLP frameworks into production.
AIRAGLangChainPower BINLPAzure AIPython
AI Research Intern
University of Essex
Sep 2023 – Aug 2024
Remote
Conducted NLP and sentiment analysis research for early BDD detection, achieving 12% F1-score improvement through transformer fine-tuning on ~10K annotated clinical text samples.
Automated end-to-end text preprocessing, feature extraction, and model evaluation workflows, reducing experiment iteration time by ~40%.
Published findings in IEEE Xplore (2024) via CNN-based sign language recognition project.
NLPTransformersCNNResearchIEEE Published
Global Technical Support Engineer
Ciena
Jan 2024 – Jun 2024
Gurugram, India
Performed data-driven diagnostics on optical networks for 15+ global telecom clients (99.5% uptime SLAs); authored automation scripts reducing case resolution time by ~25%.
Managed 50+ monthly escalations with cross-functional coordination across global support teams.
NetworkingAutomationData Analysis
Salesforce Developer & Administrator
Internship
May 2023 – Dec 2023
Remote
Automated CRM workflows with Flow Builder, reducing manual workload by 40% across 50+ users; built Lightning components, backend validation rules, and analytics dashboards.
Developed Apex triggers to extend platform functionality and managed permission sets and data integrity.
SalesforceApexLightningCRMAutomation
Frontend Developer
Ulavi Technologies PTE. Ltd
Jan 2023 – Apr 2023
Singapore (Remote)
Designed responsive frontends for travel platforms, improving user experience and performance by 15%.
Collaborated in iterative feature rollout using Agile and version control best practices.
FrontendJavaScriptResponsive DesignAgile
Research4 publications
JournalPublished July 2026
A Pilot Study of a Gamified CBT-Based Digital Approach for Body Image Distress and Appearance-Related Concerns
Human Behavior and Emerging Technologies (Wiley/Hindawi)
Co-developed a pilot-scale serious game integrating CBT techniques with AI-driven analysis to support body image distress and appearance-related concerns. The AI model interpreting user input reached 97.15% accuracy on held-out test data, outperforming a BERT-based baseline, with preliminary results pointing to gamified digital interventions as a scalable, accessible complement to traditional CBT.
97.15%
AI model accuracy interpreting user input on held-out test data
Authored chapter investigating the collaborative integration of Blockchain, the Metaverse, and IoT toward decentralized smart urban ecosystems. Explores resource optimization, citizen experience, security, and sustainability — with focus on data privacy via blockchain and real-time urban service management via IoT.
Case Study on Ethical AI-Based Decision-Making in E-Commerce Industrial Sector: Insights on McDonald's and Deliveroo
Wiley — Ethical Decision-Making Using Artificial Intelligence: Challenges, Solutions and Applications (Chapter 12)
Anushka Singh · Naman Tyagi · Dolly Sharma
Explored real-world implementation of AutoML principles at McDonald's, which optimizes food menus based on user preferences, and Deliveroo, which improved delivery times and service quality through AutoML. Examined AutoML's impact on business workflows alongside the ethical considerations organizations should weigh when adopting it.
Introduced a CNN-based model for American Sign Language recognition achieving ~99% accuracy using color + skeleton-mapped image datasets. Demonstrated significant impact of preprocessing strategy on model accuracy across varied backgrounds.
AI/ML engineer with an MS in Computer Science from NYU Tandon building production LLM applications, RAG pipelines, and agentic AI systems. Published researcher with IEEE (2024), Wiley (2025), Cambridge Scholars Publishing (2025), and Human Behavior and Emerging Technologies (2026). 6+ internships across AI, cloud, and enterprise software — I bring ideas from research to production.
Brooklyn, NY
Education
Master of Science — Computer Science
NYU Tandon School of Engineering
Aug 2024 – May 2026
B.Tech — Computer Science & Engineering (Hons. in AI & ML)
Amity University Noida
Jul 2020 – Jun 2024
Certifications
Google Cybersecurity Program Certificate · Big Data & Machine Learning Fundamentals · AWS Cloud Completion Badge · GCP Cloud Completion Badge · Azure Cloud Completion Badge · Salesforce Administrator & Developer
Away from the keyboard
F1 on race weekends, badminton and tennis, the gym, gaming, chess, always hunting for a new restaurant, and planning the next trip.
