Hi, I'm Jay Daftari.

A Pythonist

Self-driven, quick starter, passionate programmer with a curious mind who enjoys solving complex and challenging real-world problems.

Scroll

Experience

Lina Law logo

Lina Law

Forward Deployed Engineer
  • Architected an MCP server with 60+ tools to read client emails, access case files, draft and review contracts using the firm’s own Word templates, cutting turnaround from weeks to hours.
  • Orchestrated Claude-managed agent integration with a real-time collaborative legal document editor supporting concurrency for drafting, editing, and review, used daily by the firm’s lawyers.
  • Led a Slack integration automating client intake, quoting, and initial drafts with AI agents, enabling the firm to handle high-volume requests without adding headcount while maintaining lawyer approval for every deliverable.
  • Built agent observability tooling that replays each AI run step by step (its reasoning and every tool call), making every AI action auditable by non-technical teammates.
  • Established CI eval and regression gates for 380+ lawyer-built AI agents, automating 14 checks for legal-source freshness, tool permissions, and template integrity on every pull request, blocking production deployments until all checks pass.
  • Engineered a GDPR-aligned PII redaction pipeline integrating rule-based detection, LLM-assisted redaction, lawyer-in-the-loop review, and pre-inference pseudonymization, enabling secure AI adoption for confidential legal matters.
  • Tools: Python, TypeScript, Next.js/React, PostgreSQL (Supabase), Claude, MCP
May 2026 - Oct 2026 | Remote
Wequity logo

Wequity

AI Engineer
  • Built Nora’s Smart Drafting engine on Claude Opus 4.6 (GPT-5.4 and Gemini 2.5 Pro fallbacks), auto-filling legal templates from client files with a cited source for every field, so lawyers verify answers instead of re-reading documents.
  • Hardened the drafting pipeline (FastAPI, Cloud Run) with parallel extraction, per-file failure isolation and Slack alerts, so one bad upload no longer fails an entire draft.
  • Cut first paint by 80% (12.2s to 2.4s) and raised Nora’s Lighthouse performance score from 55 to 68 by reworking startup loading; fixed Postgres bloat by moving to UUIDv7 keys and dropping 34 unused indexes.
  • Tools: Python, FastAPI, TypeScript, React, PostgreSQL (Supabase), GCP Cloud Run, Docker
Feb 2026 - May 2026 | Remote
Vicktoria logo

Vicktoria AI

Founding Engineer
  • Launched vicka.ai, a chat and voice AI coach, deploying a fine-tuned Mixture-of-Experts model with vLLM on RunPod GPUs, activating only 4B parameters per token and eliminating per-token API fees through self-hosted inference.
  • Led the development of a retrieval-augmented generation (RAG) pipeline grounding responses in the founder’s methodology using BGE embeddings and pgvector, with zero-downtime re-indexing to maintain source availability during updates.
  • Engineered a hands-free voice pipeline integrating speech-to-text, LLM inference, and self-hosted Kokoro TTS (Python, FastAPI), reducing perceived latency through sentence-level streaming playback.
  • Integrated real-time camera and photo inputs for multimodal AI interactions, implementing browser-side image resizing to 768×768 for vision model inference.
  • Automated CI/CD with GitHub Actions, building Docker images and deploying to AWS EC2 on every merge to main via OIDC and SSM, eliminating SSH keys and stored AWS credentials.
  • Prototyped a talking-avatar service on Modal H100 GPUs with scale-to-zero inference, eliminating idle compute costs and achieving an estimated $0.05–$0.20 per reply.
  • Tools: Node.js, React, Python, FastAPI, LangChain, PostgreSQL, Docker
Sep 2025 - Apr 2026 | New York City, New York
DEP logo

NYC DEP

Summer Graduate Intern
  • Re-engineered public-sector enterprise applications supporting contracts, procurement, and payment systems handling over $3 billion in annual transactions using Web Services, jQuery, ASP.NET, and VB.NET.
  • Utilized MS SQL Server, Python, Excel, Power BI, DAX Studio, and Report Viewer for end-to-end ETL processes, data analysis, and enterprise reporting.
  • Tools: SQL, Power BI, Python
Jun 2025 - Aug 2025 | New York City, New York
NYU logo

New York University

Course Assistant
  • Tools: Git, C/C++, python
  • Managed Opensource course development and assignment guidelines in Python and C++ with Prof. Kamen Yotov.
  • Reviewed code for integrating an AI conversational bot with mail clients as a reusable package on GitHub, graded HW for over 50 students, and enforced best practices through continuous integration (CI/CD) and automated testing.
Jan 2025 - May 2025 | New York, USA
IDFC logo

IDFC First Bank

Application Engineer
  • Contributed to core product development of tablet and web apps impacting 20M customers using Agile and React.js.
  • Implemented UPI, VKYC, and other features in the front end across pre- and post-account opening journeys.
  • Collaborated with cross-functional teams to design and document API contracts for RESTful services using Java Micronaut and Kafka, resulting in a 10% increase in online zero-balance account openings.
  • Optimized product quality by increasing unit test coverage to 97% with Jest, conducting browser testing, and boosting Lighthouse performance score from 70% to 85%.
  • Configured Grafana and Prometheus dashboards to monitor performance metrics, reducing issue detection time by 30% and providing detailed impact reports to the business.
  • Tools: React Js, React Native, GoCD, Micronaut, Jest, Grafana, Prometheus, kubernetes, Docker
Mar 2022 - Aug 2024 | Bengaluru, India
Vit logo

Center of Advanced Data Science

Research Intern
  • Spearheaded research on advanced machine learning models for water quality assessment, achieving 90% accuracy, and utilized LIME and SHAP for model interpretability.
  • Tools: Python, Keras,Scikit-learn, Flask, JavaScript
Jul 2021 - Aug 2021 | Chennai, India
Premware logo

Premware Services India LLP

MEAN Stack Developer
  • Led the development of Surati Taxi’s scalable backend with REST API, cutting customer request response time by 50%
  • Implemented web sockets, notifications, and real-time location tracking, enhancing user engagement by 40% and reducing response latency by 20%.
  • Tools:Node JS, Angular Js, Mongo DB, Firebase
Feb 2021 - Mar 2021 | Surat, India

Hackathons

Flight log · to

  1. Finalist, at Cornell AI Hackathon

    Top 6 teams

  2. Winner, Human Impact Track, at NYC Spark Hack

    NVIDIA, Antler and Acer

    , New York, USA

  3. 3rd Place, at GIDE Hackathon @ NYC Tech Week

    Part of the a16z ecosystem

    , New York, USA

Open Source

selimfirat/pysad

1 merged PR
  • Streaming anomaly detection in Python: 16 online detectors plus calibration and prequential evaluation for real-time data (300+ GitHub stars).
  • #408: Completed the type annotations of pysad.transform.ensemble (issue #376), including the abstract bucket-combination method, and added the subpackage to mypy's enforced list so it stays typed.
  • Tools: Python, mypy
Oct 2026

lastmile-ai/mcp-agent

2 merged PRs
  • Build effective agents using Model Context Protocol and simple workflow patterns (8.5k+ GitHub stars).
  • #608: Fixed slow feedback on uv run mcp-agent init (issue #538) by adding a lightweight bootstrap entry point that shows a progress indicator while heavy imports load.
  • #613: Follow-up so the loading spinner only appears in interactive terminals, keeping CI and other non-interactive runs clean.
  • Tools: Python, CLI, uv
Nov 2025

hart-v2

2 merged PRs
  • Hybrid Autoregressive Transformer (HART) image-generation pipeline.
  • #1: Built the working end-to-end pipeline.
  • #2: Added batch latency profiling to measure and speed up batch generation.
  • Tools: Python, PyTorch
Oct 2025

1,018 contributions in 2026

Projects

Mission manifest

  1. 01 · Web apps ·

    1Password for your AI memory: Rust engine and Tauri desktop app that indexes a user's email, chats and files on their own device, then serves that context to Claude, ChatGPT and other AI tools over MCP.

    Highlights

    • A Rust core that imports 45 exports and live-syncs Gmail, Google Drive, Calendar, Spotify, GitHub and Telegram into plain files, searched with ripgrep's engine and no index to build.
    • An MCP server with 7 read-only tools for Claude Desktop, Cursor and ChatGPT or Claude connectors; each app sees only the folders it is allowed, and every request is logged.
    • A Tauri desktop app with on-device text from photos and scans (Apple Vision), photo search by description with EmbeddingGemma, and an AES-256 encrypted drive on macOS.
    • A launch site on Vercel with a 3D product story, a waitlist and the hosted investor deck.

    Visit site savant-green.vercel.app (opens in a new tab)

  2. 02 · Web apps

    An editorial magazine platform with a writers' desk and an editor review workflow.

    Highlights

    • Writers draft articles in a rich-text editor with images, links and YouTube embeds, then submit them for review.
    • Built an editorial desk where editors approve or reject submissions, assign homepage placements and manage sponsor bands.
    • Newsletter sign-ups and editor email notifications through Resend, media stored on S3, plus an SEO sitemap and llms.txt.

    Visit site shelclub.vercel.app (opens in a new tab)

  3. 03 · ML & data

    An AI tool that watches live NYC traffic cameras and forecasts street risk.

    Highlights

    • Won the Human Impact Track at NYC Spark Hack (NVIDIA × Antler × Acer) with a working prototype built in one weekend.
    • Designed the app's screens and put 364 live NYC DOT traffic cameras on a map, each analyzed on demand by a local Gemma 4 vision model that flags hazards.
    • Built a RAPIDS pipeline on an NVIDIA DGX Spark that maps NYC Open Data onto a hexagon grid and trains an XGBoost model to forecast risk, shown as a heatmap with a time-of-week slider.

    View source jaydaftari/person-of-interest (opens in a new tab)

  4. 04 · ML & data

    A date planner that learns from how you rated past dates and recommends your next one.

    Highlights

    • Built at Hacklytics 2026 with a team of four. An LLM (Llama 3.1 on Groq) scores each date, described in plain words, on 9 dimensions from cost and energy to romance and conversation depth.
    • The best-rated dates become a preference vector, matched against 200 date activities in the Actian vector database by cosine similarity to suggest the next two.
    • Built the React front end (the dashboard, sign-in page, recommendation flow, and date history and analytics views) and containerized the app with Docker Compose.

    View on Devpost devpost.com/software/mynextdate (opens in a new tab)

  5. 05 · ML & data

    A self-learning ad-blocking proxy that classifies new domains with an LLM and remembers every verdict.

    Highlights

    • Co-built a proxy (mitmproxy) that intercepts HTTP and HTTPS traffic, lets an unseen domain through at once and classifies it in the background as an ad or legitimate with an LLM (Llama 3.1 8B on Groq).
    • Each verdict is cached in a trie keyed by TLD, domain pattern and subdomain pattern, so later requests are blocked or allowed in microseconds without another API call.
    • The classifier reads the domain, redirects, network metadata (headers, response size, cookies, timing) and the first 800 characters of content; a whitelist of 150+ trusted domains is never blocked.

    View source jaydaftari/LLM-Based-AdBlocker (opens in a new tab)

  6. 06 · Deep learning

    A faster HART text-to-image generation pipeline.

    Highlights

    • Scaled and optimized the HART image generation pipeline, doubling batch processing speed while maintaining FID score quality for prompt streams.
    • Built a Gradio demo that filters unsafe prompts with ShieldGemma-2B.

    View source jaydaftari/hart-v2 (opens in a new tab)

  7. 07 · Deep learning

    Jupyter Notebook containing the code

    Highlights

    • Evaluated ResNet-34’s adversarial robustness on a 500-image ImageNet subset using FGSM, PGD, and patch attacks—assessing cross-model transferability to uncover architecture-specific vulnerabilities.

    View source jaydaftari/Project3DL (opens in a new tab)

  8. 08 · Deep learning

    Gradient statistics that guide pruning and gradient prediction.

    Highlights

    • Performed statistical gradient analysis on large language models (LLM) to identify distribution trends.
    • Used those trends for weight pruning and gradient prediction, improving training efficiency for edge deployment.

    View source jaydaftari/EfficientAI_Project (opens in a new tab)

  9. 09 · Deep learning

    Jupyter Notebook containing the fine-tuning code

    Highlights

    • Achieved 98.73% test accuracy on AG News headline classification by fine-tuning a RoBERTa model with LoRA, reducing trainable parameters to enhance training efficiency.

    View source jaydaftari/DeepLearningProject2 (opens in a new tab)

  10. 10 · ML & data

    Website which helps user to remain focused.

    Highlights

    • Co-developed a real-time eye-tracking application with 90% accuracy.
    • Coded a webcam-based tracking system leveraging VLM, reducing alert latency to 4 seconds.

    View source danielkaijzer/Focus-App (opens in a new tab)

  11. 11 · ML & data

    Analytics and forecasting for 5+ years of NYC subway ridership.

    Highlights

    • Built an interactive Streamlit dashboard analyzing 5+ years of NYC MTA ridership data, processing millions of records using PySpark to identify trends, peak usage patterns, and service disruptions across multiple time scales.
    • Developed automated ETL pipelines with data quality monitoring, feature engineering, and geospatial visualizations (GeoPandas, Folium) to map station-level accessibility patterns.
    • Added 7-day ridership forecasts with Prophet and an XGBoost demand model explained with SHAP.

    View source jaydaftari/CSGY-6513-Big-Data-Finals-Project (opens in a new tab)

  12. 12 · ML & data

    Created it using K-means clustering algorithm and training resulted clusters on cluster number using KNN.

    Highlights

    • Created a page where users can enter a song and receive recommended songs.
    • Formed clusters based on various features such as popularity, danceability, energy, etc., after performing feature selection.
    • Achieved 98% accuracy using the K-Nearest Neighbors (KNN) algorithm.
    • Utilized FuzzyWuzzy to validate song names.

    View source jaydaftari/music_recommendation_system (opens in a new tab)

  13. 13 · C & C++

    Generic Lock/unlock pattern recorder using STM32F429I

    Highlights

    • Register Lock pattern using different complicated hand patterns.
    • Used same lock pattern to unlock and different pattern to get error.

    View source jaydaftari/Embedded_Sentry (opens in a new tab)

  14. 14 · ML & data

    A website that detects hands gesture and print it.

    Highlights

    • Recognizes Indian Sign Language hand signs with 90% test accuracy and prints them.
    • Trained on different models such as: Naive Bayes K-Nearest Neighbours, Support Vector Machines, Convolution Neaural Network, etc. and selected best model with low latency(SVM)

    View source jaydaftari/Dexterity (opens in a new tab)

  15. 15 · Web apps

    Online Scholarship Portal for students

    Highlights

    • Engineered a website that helps users find 200+ scholarships they are eligible for in a few clicks.
    • Integrated a 24/7 chatbot, reducing query resolution time by 4 minutes per interaction.

    View source jaydaftari/Sportal (opens in a new tab)

  16. 16 · C & C++

    A C++ program that gives dishes based on given ingrediants .

    Highlights

    • Architected a 2-tier system using graph and hash table data structures to suggest personalized recipes from user-input ingredients.
    • Added admin features for managing the recipe database, reducing data entry time by 40%.

    View source jaydaftari/cluelesschef (opens in a new tab)

  17. 17 · Web apps

    Website that detects face.

    Highlights

    • Developed a website which detects face .
    • Has login and signup functionality

    View source jaydaftari/faceapp (opens in a new tab)

Publications

Explainable AI Framework for Multi-label Classification using Supervised Machine Learning Models

Water is known as a "universal solvent" as it is extraordinarily frail against contamination. Water quality standards are developed based on logical evidence of the effects of hazardous compounds on a certain quantity of water used. Classification techniques of machine learning can be employed to understand the water quality status. In this work, supervised machine learning models are being implemented to classify water quality indexes, and the Smote analysis is used to handle the imbalance in the dataset. An artificial neural network model is built using the features such as Oxygen, pH, temperature, total suspended sediment, turbidity, nitrogen, and phosphorus as inputs and water quality check as the target variable. This target variable is created using Canadian Council of Ministers of the Environment Water Quality Index, and the model works with an accuracy of 87%. The classification is done on XGBoost model as well and it performs with an accuracy of 90%. The explanations for predictions of these models for a data instance were performed using explainable artificial intelligence tools such as LIME and SHAP. The results and interpretations for the predictions seem to be more promising and attractive making the proposed models more interpretable, accurate and efficient. Through our research, we can benefit our readers by providing them clarity about exactly what features are having more influence on water quality than others from different machine learning algorithms. This will help the developers to gain insights into the significant factors of poor water quality and how to overcome that.

Explainable AI Framework for Multi-label Classification using Supervised Machine Learning Models | Nov 2022

Efficient Dynamic Phishing Safeguard System Using Neural Boost Phishing Protections

The instances of privacy and security have reached the point where they cannot be ignored. There has been a rise in data breaches and fraud, particularly in banks, healthcare, and government sectors. In today’s world, many organizations offer their security specialists bug report programs that help them find flaws in their applications. The breach of data on its own does not necessarily constitute a threat or attack. Cyber-attacks allow cyberpunks to gain access to machines and networks and steal financial data and esoteric information as a result of a data breach. In this context, this paper proposes an innovative approach to help users to avoid online subterfuge by implementing a Dynamic Phishing Safeguard System (DPSS) using neural boost phishing protection algorithm that focuses on phishing, fraud, and optimizes the problem of data breaches. Dynamic phishing safeguard utilizes 30 different features to predict whether or not a website is a phishing website. In addition, the neural boost phishing protection algorithm uses an Anti-Phishing Neural Algorithm (APNA) and an Anti-Phishing Boosting Algorithm (APBA) to generate output that is mapped to various other components, such as IP finder, geolocation, and location mapper, in order to pinpoint the location of vulnerable sites that the user can view, which makes the system more secure. The system also offers a website blocker, and a tracker auditor to give the user the authority to control the system. Based on the results, the anti-phishing neural algorithm achieved an accuracy level of 97.10%, while the anti-phishing boosting algorithm yielded 97.82%. According to the evaluation results, dynamic phishing safeguard systems tend to perform better than other models in terms of uniform resource locator detection and security.

Electronics Journal (MDPI Publication)| Sep 2022

Skills

41 skills · 4 constellations

Matching happens in your browser; nothing is sent.

41 skills in 4 constellations. Arrow keys move between stars, Page Up and Page Down jump between constellations, Enter pins a star, Escape clears. Details appear in the chart key after the chart.

Languages

AI / ML

Data & Infra

Web & Apps

Education

New York University

New York, USA | Sep 2024 - May 2026

Degree: Master of Science in Computer Engineering
CGPA: 3.7/4.0

Relevant Coursework:

  • Efficient AI and Hardware Accelerator Design
  • Deep Learning
  • Introduction to Machine Learning
  • AI/ML for Networks
  • Big Data
  • Internet Architecture and Protocols
  • Real-Time Embedded Systems
  • Introduction to System Engineering

Vellore Institute of Technology

Chennai, India | Jul 2018 - May 2022

Degree: Bachelor of Technology in Computer Science
CGPA: 3.86/4

Relevant Coursework:

  • Data Structures and Algorithms
  • Database Management Systems
  • Operating Systems
  • Computer Networks
  • Machine Learning
  • Image Processing
  • Natural Language Processing

Contact

+19299795976

jd5829@nyu.edu

Book a call · Google Calendar

github.com/jaydaftari

linkedin.com/in/jay-daftari

x.com/jaydaftari19