Skip to content
View DebadityaHait's full-sized avatar
:shipit:
:shipit:

Sponsoring

@rclone

Block or report DebadityaHait

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
DebadityaHait/README.md

Overview

Computer Science Engineering student focused on AI systems architecture, edge-native inference optimization, and distributed storage runtimes. Active open-source contributor across major ecosystems including Google TensorFlow, PyTorch, Rust-native pnpm, Vite, Deno, AWS Lambda, Node Redis, CERN TIGRE, and Cloudflare.

Previous research includes internships at Samsung R&D Institute India (lightweight encoder-only guardrails for AI safety, ModernBERT-large 35ms inference) and IIT Madras (Physics-Informed Neural Networks for Alzheimer's progression modeling). Holds 9x Microsoft Certifications, Claude Certified Architect, and competes in Division 1 competitive programming (LeetCode Guardian Top 2%, Codeforces Expert, CodeChef 5-Star).


Open Source PRs Merged Upstream Competitive Rating Certifications Education
59 Pull Requests
(47 Unique Codebases)
16 Merged PRs
(TensorFlow, PyTorch, pnpm, Vite)
LeetCode Guardian (Top 2%)
Codeforces ExpertCodeChef 5★
9x Microsoft Certified
Claude Certified Architect
BTech CSE @ SRMIST
BS Data Science @ IIT Madras

Open Source Contributions

Summary: 59 pull requests across 47 unique repositories (16 merged upstream, 25 active open) across Rust, C/C++, Python, Go, and TypeScript.

Total PRs Merged PRs Open PRs Languages

tensorflow stars pytorch stars kubernetes stars nextjs stars electron stars deno stars vite stars pnpm stars neovim stars posthog stars uv stars rclone stars


Merged Upstream Pull Requests

Repository Pull Request Stack / Diff Technical Solution & Impact
Rust-Native pnpm Engine
pnpm/pnpm
pnpm stars
#13059

+1,033 / -125
Git Resolver Manifest Reader & Security Hardening: Implemented a resolve-time archive fetch and manifest reader in Rust (GitResolver) for pacquet, fixing a lockfile generation panic during git-hosted dependency installs. Patched option-injection vulnerability in git clone (--upload-pack) and prevented subpath directory traversals (#path:/../..) across 5 crates.
Rust-Native pnpm Engine
pnpm/pnpm
pnpm stars
#13056

+260 / -12
Git Specifier Normalization: Developed a cycle-safe Git specifier normalization module in Rust, preventing false-stale lockfile errors on equivalent protocols (git+https:// vs git://) and hosted shortcuts (github:) during frozen installs.
Google TensorFlow Core
tensorflow/tensorflow
tensorflow stars
#124961

+28 / -9
oneDNN Layout Rewriting Crash Fix: Preserved 64-bit integer convolution attributes across oneDNN layout rewriting passes in TensorFlow's C++ kernel engine, converting fatal process aborts into clean InvalidArgumentError exceptions.
PyTorch Distributed Core
pytorch/pytorch
pytorch stars
#191831
30731ee

+4 / -2
Symmetric-Memory Triton Hooks: Resolved Ruff B018 static analysis warnings in torch.distributed._symmetric_memory while preserving deliberate lazy module attribute lookup mechanics across NVSHMEM and Shared Memory Triton utilities.
Vite Build Engine
vitejs/vite
vite stars
#22947

+100 / -2
Chunk Import Map CSS Caching: Fixed browser cache collisions in Vite's experimental chunk import-map engine by mapping dynamically extracted CSS assets back to parent Rollup chunks and generating stable Web Import Map specifiers.
PostHog CLI
PostHog/posthog
posthog stars
#70314

+248 / -30
Sourcemap Upload Concurrency: Replaced a hardcoded global Rayon thread pool with an isolated, configurable concurrency pool in the Rust PostHog CLI, exposing --concurrency controls to prevent CI/CD network and resource exhaustion.
Node Redis
redis/node-redis
node-redis stars
#3388

+66 / -8
Sentinel Outage Retry Bounding: Enforced maxCommandRediscovers bounds on post-connect Sentinel topology rediscoveries during cluster outages and implemented #resetInBackground() for graceful promise rejections.
rclone Cloud Sync
rclone/rclone
rclone stars
#9712

+100 / -55
Context-Aware SDK Calls: Refactored Dropbox backend to propagate caller context.Context down to HTTP execution sites, eliminating leaked goroutines and ensuring prompt cancellation during interrupted transfers.
CERN TIGRE
CERN/TIGRE
TIGRE stars
#762

+18 / -2
Tomography Calibration Math: Corrected Varian CBCT rotational scan direction math (CW vs CC) and uncompressed XIM pixel buffer decoding in CERN's TIGRE reconstruction toolbox, verified against Zenodo phantom datasets.
Deepset Haystack AI
deepset-ai/haystack-core-integrations
haystack stars
#3543

+141 / -3
SentenceTransformers Quantization: Resolved divide-by-zero errors in single-text int8/uint8 embeddings by introducing static pre-calibrated quantization ranges (shape: (2, D)).
IBM / DS3 Docling
docling-project/docling
docling stars
#3875

+45 / -2
Document Pipeline Crash Fix: Added defensive page validation in SimplePipeline to skip image enrichment for PPTX native charts without page renders, fixing an index out-of-bounds conversion crash.
Neovim Core
neovim/neovim
neovim stars
#41067

+22 / -2
UI2 Highlight Retention: Engineered buffer-scoped highlight lifecycle checks in Neovim's C command-line UI subsystem to preserve substitution confirmation highlights under nohlsearch.
Twenty CRM
twentyhq/twenty
twenty stars
#23945

+80 / -2
Relational Data Preservation: Patched entity deduplication in NestJS/GraphQL backend services, preserving polymorphic relations and timeline activity records during contact merges.
AWS Lambda Powertools
aws-powertools/powertools-lambda-python
powertools stars
#8374

+50 / -5
OpenAPI Parameter Deduplication: Omitted redundant Content-Type header parameters from generated OpenAPI specifications while preserving runtime media-type validation.
Cloudflare Workers SDK
cloudflare/workers-sdk
workers-sdk stars
#14668

+140 / -1
Local Explorer UI Containment: Added responsive vertical scroll containment to the worker selection UI, resolving clipping bugs for multi-worker environments.
Apache DataFusion Comet
apache/datafusion-comet
comet stars
#5458

+11 / -0
Performance Tuning Guide: Documented columnar-to-row (C2R) transition overhead and stage-revert tuning strategies for wide and nested Spark schemas.
MUI Material-UI
mui/material-ui
material-ui stars
#48883

+32 / -3
Slot Prop Inheritance: Documented React Autocomplete slot prop inheritance rules to prevent rendering bugs during multi-select custom input adornment configuration.

Selected Active Open Pull Requests

Repository PR Link Stack Technical Focus
Apache DataFusion Ballista
apache/datafusion-ballista
Ballista stars
#2397 Rust utoipa +523 / -26 OpenAPI v3 specification generator and JSON endpoint for distributed scheduler.
Delta Lake Core
delta-io/delta-rs
delta-rs stars
#4685 Rust Parquet +43 / -4 Nested struct field override preservation for DataFusion Parquet leaf pruning.
Hugging Face Safetensors
huggingface/safetensors
safetensors stars
#826 Rust Python +34 / -3 Strict memory-layout validation rejecting non-C-contiguous NumPy arrays.
Kubernetes Core
kubernetes/kubernetes
kubernetes stars
#140791 Go Kubelet +57 / -4 Pre-allocation bypass for zero-valued device requests in Kubelet Device Manager.
Deno Runtime Core
denoland/deno
deno stars
#36532 Rust JS +52 / -8 Teardown socket cleanup closing accepted idle HTTP sockets in Node compatibility layer.
Google TensorFlow Core
tensorflow/tensorflow
tensorflow stars
#124960 C++ Runtime +9 / -1 Inter-op compute pool handling for negative thread configuration values.
Google TensorFlow Core
tensorflow/tensorflow
tensorflow stars
#124959 C++ oneDNN +35 / -1 Rewrite error propagation preventing aborts on invalid quantized pooling attributes.
PyTorch Deep Learning
pytorch/pytorch
pytorch stars
#191876 Python torch.compile +65 / -3 Input shape validation for pixel_shuffle meta-registrations and compile paths.
Astral uv
astral-sh/uv
uv stars
#20855 Rust +115 / -27 Staged atomic binary updates on Windows eliminating file-lock permission panics.
Apple CoreMLTools
apple/coremltools
coremltools stars
#2767 Python MIL +46 / -5 MIL compiler validation rejecting symbolic classifier probability shapes.
Electron Core
electron/electron
electron stars
#52737 C++ TS +37 / -9 BaseWindow normalization ensuring compliance with Partial<Rectangle> bounds.
Vercel Next.js
vercel/next.js
nextjs stars
#96111 TypeScript +40 / -1 Middleware rewrite search parameter retention in App Router request URLs.
AWS Lambda Rust Runtime
aws/aws-lambda-rust-runtime
lambda-rust stars
#1162 Rust SNS +45 / -2 Fixed millisecond timestamp precision for SNS signature verification.
Ansible Automation
ansible/ansible
ansible stars
#87326 Python +224 / -4 Direct tarball download command in ansible-galaxy role for air-gapped systems.
Graphify Knowledge Graph
Graphify-Labs/graphify
graphify stars
#2728 Python +176 / -51 Windows CLI candidate probing preventing PATH-shadowing subprocess errors.
Graphify Knowledge Graph
Graphify-Labs/graphify
graphify stars
#2726 Python +85 / -17 Fail-closed artifact preservation during community labeling failures.
Storybook Ecosystem
storybookjs/storybook
storybook stars
#35520 TypeScript +250 / -10 Angular Signal Query isolation preventing property overwrites in decorators.
Microsoft PowerToys
microsoft/PowerToys
PowerToys stars
#49375 C++ +56 / -10 Culture-aware digit formatting for Command Palette calculator results.
Pytest Framework
pytest-dev/pytest
pytest stars
#14707 Python +69 / -13 Custom TOML file root table parsing for pytest -c executions.
Node Version Manager
nvm-sh/nvm
nvm stars
#3891 Shell +45 / -0 Shell trap restoration for Zsh extendedglob terminal settings.
Nx Monorepo Toolkit
nrwl/nx
nx stars
#36347 TypeScript +409 / -452 Compiler plugin migration to minimizer-webpack-plugin with cssnano.
NousResearch Hermes Agent
NousResearch/hermes-agent
hermes stars
#70711 Python +63 / -2 Local slash command auto-completion support in terminal interface.
NousResearch Hermes Agent
NousResearch/hermes-agent
hermes stars
#70381 TypeScript +175 / -1 Reactive state refresh hooks for active chat sessions in dashboard UI.
Traceroot AI
traceroot-ai/traceroot
traceroot stars
#1860 TypeScript +28 / -12 Ambiguous filter field glyph replacement with neutral fallback icons.

Experience & Research

Samsung R&D Institute India

PRISM Research Intern (AI Safety & Systems Architecture)  |  April 2025 – September 2025

  • Encoder-Only Guardrails: Fine-tuned ModernBERT-large (395M parameters) into a lightweight classifier for prompt injection and jailbreak detection across 75k+ augmented adversarial samples (InjectGuard, BIPIA, NotInject, WildGuard-Benign, PINT).
  • Latency & Benchmarks: Reached 92.70% accuracy and 0.9021 F1-score on a 315-prompt adversarial benchmark with 35 ms inference latency (~300x faster than generation-based alternatives), outperforming NVIDIA NeMo Guardrails, Llama Guard 4, Prompt Guard, and DeBERTa-v3.
  • Publication: Co-authored "AI Guardrails: Performance-Optimized Guardrail System for Secure LLM and Agentic AI Interactions", Accepted to ICSCCC-2026.

Indian Institute of Technology Madras

AI Research & Prototyping Intern (Scientific ML & PINNs)  |  June 2025 – August 2025

  • Physics-Informed Neural Networks (PINNs): Embedded biological differential equations directly into PyTorch loss functions for spatiotemporal modeling of Alzheimer's disease progression on longitudinal MRI data.
  • Accuracy Gains: Implemented automatic differentiation for PDE residuals, collocation sampling, and adaptive loss weighting, delivering a 27% improvement in early neurodegeneration prediction accuracy over conventional deep learning baselines.

Featured Projects

S3-Compatible Edge Distributed Object Storage

  • Architected edge S3 object storage achieving 95% cost reduction via multi-tier caching (Edge Cache hot reads, R2 warm cache, Telegram Bot cold storage).
  • Sub-50ms metadata queries via Cloudflare D1 with SHA-256 and Content-MD5 integrity checks.
  • 100% compatibility with S3 CLI/SDKs via reverse-engineered S3 XML response serialization.

Serverless Edge Observability & Incident Intelligence

  • Serverless edge uptime monitoring maintaining 99.99% uptime across distributed endpoints.
  • Decoupled 1-min health checks from AI anomaly analysis via Cloudflare Queues for zero latency impact.
  • Integrated Workers AI (LLaMA-3.1-8B) for automated root-cause summaries, reducing MTTR by 70%.

Hyperlocal Quick-Commerce Price Intelligence

  • Aggregated live inventory & pricing across Zepto, Blinkit, Instamart, and Flipkart Minutes with 85% discovery speedup.
  • Reverse-engineered marketplace protocols with custom Node scrapers maintaining 99.8% extraction accuracy.
  • Mapped 10k+ SKUs into unified groups via token similarity algorithms; scheduled 5,000+ daily alert checks.

Edge AI Clinical Diagnostics System

  • Hybrid MiT-b3 + CNN + P-scSE segmentation achieving 92.33% classification accuracy and 87.64% Dice.
  • Quantized 16-class EfficientNet-B0 to TFLite (48 MB → 12.5 MB) for low-end mobile SoCs (Snapdragon 439).
  • Validated in 28-day clinical pilot across 2 Tamil Nadu rural clinics (4.7% area error vs calipers). Submitted to MIDL 2026.

Distributed Metadata-First Cloud Storage Engine

  • Separated Postgres metadata from binary chunks stored via Telegram MTProto API, achieving 81% storage cost savings.
  • Adaptive 20 MB chunking with SHA-256 integrity; handled 1,240+ operations over 6-week beta with zero data loss.
  • Server-signed HMAC-SHA256 request authorization with sub-42ms metadata API responses.

Vernacular & Non-Native English ASR Pipeline

  • Fine-tuned Whisper on 180+ hrs Indian English and 62 hrs Hinglish audio, yielding a 22% reduction in Word Error Rate.
  • 91% accuracy in detecting disfluencies (fillers, backtracking) with sub-280ms CPU inference across 650+ practice sessions.
  • Deployed to 47 non-native speakers, delivering 37% average reduction in filler words over 3 months.

Competitive Programming & Hackathons

Platform / Event Distinction Standing
LeetCode Guardian Top 2% Globally
Codeforces Expert Peak Candidate Master
CodeChef 5-Star Coder Division 1 Competitor
Goldman Sachs Hackathon 2026 All-India Rank 42 (Quant Track) Top ~0.3% of ~14,000 participants
Goldman Sachs Hackathon 2026 All-India Rank 144 (CS Track) Top ~0.9% of 15,953 participants
Amazon ML Summer School 2026 Selected Scholar Top 2.2% (Top 3,000 of 134,000+ applicants)
Flipkart GRiD 8.0 (2026) National Semifinalist (Software Dev) Top ~2.3% of 130,000+ applicants
Flipkart GRiD 7.0 (2025) National Semifinalist (Software Dev) Selected among 160,000+ participants
Adobe India Hackathon 2025 National Semifinalist (Creator of Outlinr) Selected from 260,000 registrations
Founder’s Scholarship Full Academic Merit Scholarship SRM Institute of Science and Technology (2023)

Certifications

Issuer Credential Title Verification ID
Microsoft Certified: Fabric Data Engineer Associate 2CFE44AF9A03CA00
Microsoft Certified: Azure Data Scientist Associate 6B609B2E62C7DD1B
Microsoft Certified: Azure Developer Associate 7C24709582312829
Microsoft Certified: Azure AI Engineer Associate C6A78B398CD5D2CE
Microsoft Certified: Fabric Analytics Engineer Associate 31C332B22EA60C45
Microsoft Certified: Power BI Data Analyst Associate AB65FD5A69BEE6C9
Microsoft Certified: Azure Fundamentals C98BA2EF3EA717B0
Microsoft Certified: Azure Data Fundamentals A05553E0E826BF55
Microsoft Certified: Azure AI Fundamentals 743A61D9C9700E53
Claude Certified Architect - Foundations Anthropic Verified
OCI Certified AI Foundations Associate Oracle Cloud Verified
Machine Learning Specialization Stanford / DeepLearning.AI

Technical Skills

========================================================================================================
  CORE LANGUAGES       ::  Rust  •  C / C++  •  Python  •  TypeScript  •  Go  •  SQL  •  JavaScript
  SYSTEMS & RUNTIMES   ::  Deno  •  Node.js  •  Linux Kernel APIs  •  oneDNN  •  Triton  •  Electron
  AI / ML / TENSORS    ::  PyTorch  •  TensorFlow  •  TFLite  •  Hugging Face  •  Safetensors  •  PINNs
  EDGE & CLOUDFLARE    ::  Cloudflare Workers  •  D1  •  R2  •  KV  •  Queues  •  Workers AI  •  Pages
  BACKEND & STORAGE    ::  FastAPI  •  Next.js App Router  •  PostgreSQL / Neon  •  Prisma  •  Redis  •  S3
  DEV TOOLS & DEVOPS   ::  Git  •  Docker  •  Wrangler  •  Rollup / Vite  •  pnpm  •  uv  •  Pytest  •  Vitest
========================================================================================================


Education

  • SRM Institute of Science and Technology, Kattankulathur, Tamil Nadu
    BTech in Computer Science and Engineering (2023 – Present)
    Focus: Computer Architecture, Operating Systems, Data Structures & Algorithms, Distributed Systems, AI Systems Architecture.

  • Indian Institute of Technology Madras
    BS in Data Science and Applications (2024 – Present)


Contact

LinkedIn Connect   Send Email   GitHub Follow

Pinned Loading

  1. WoundDoc WoundDoc Public

    AI-Powered Chronic Wound Monitoring App for Clinicians | Built as my 6th Semester Minor Project. Features 16-class wound classification, 30+ tissue segmentation (DSC 87.64%), perspective-corrected …

    TypeScript

  2. deepxde-alzheimer-aggregation-pinn deepxde-alzheimer-aggregation-pinn Public

    DeepXDE implementation of physics-informed neural networks to identify optimal reaction kinetic models for Aβ peptide aggregation pathways.

    HTML

  3. outlinr outlinr Public

    Intelligent Document Structure Extractor - Adobe India Hackathon 2025 solution for automated PDF outline extraction using hybrid ML approach with PyMuPDF and Random Forest classifier

    Python

  4. PomoTogether PomoTogether Public

    A collaborative Pomodoro timer app built with React Native that allows users to create and join focused work sessions with friends, colleagues, or study groups.

    TypeScript 1

  5. OnyxBin OnyxBin Public

    OnyxBin is a self-hosted cloud storage and file sharing platform built with Next.js and FastAPI. It uniquely leverages Telegram as a binary object store, Neon Postgres for metadata, and Stack Auth …

    Python