╔══════════════════════════════════════════════════════════════════════════╗ ║ AYUSHMAN@INFOSAGE05 ~ AI/ML Engineer ~ IIT Kharagpur ║ ╠══════════════════════════════════════════════════════════════════════════╣ ║ $ whoami ║ ║ > B.Tech. (Hons.) Mfg. Science + M.Tech. Industrial Engg. & Management ║ ║ > Micro Spl. AI & Applications · CGPA 8.41 · IIT Kharagpur ║ ║ > Building GenAI systems, multi-agent swarms, and competitive ML. ║ ║ > Active contributor to vLLM, Unsloth, and the LLM serving ecosystem. ║ ╚══════════════════════════════════════════════════════════════════════════╝
$ neofetch --info ayushman
OS : IIT Kharagpur · Dual Degree (Mfg. Science) · Micro Spl. AI & Applications
Kernel : AI/ML Engineering, GenAI, Multi-Agent Systems
Uptime : 2023 – Present
Shell : Python · C/C++ · TypeScript · SQL
WM : PyTorch · LangGraph · FastAPI · Docker
Resolution : 8.41 CGPA · Competitive ML · OSS Contributor
Location : Kharagpur, India
Theme : Dark-terminal-greenThese are the strongest original projects I’ve shipped. Click a repo to see the source.
| # | Repo | Language | What it does |
|---|---|---|---|
| 1 | qwen-swarm | Python | RepoPilot — zero-copy multi-agent software engineering system. SGLang RadixAttention cut latency by 3,500ms and hit 83.3% cache hit rate. tree-sitter + NetworkX repo intelligence, Docker sandbox validation, Modal GPU inference. Built for Qwen Cloud Global AI Hackathon. |
| 2 | sherlock-candidate-engine | Python | Real-time candidate identification & authenticity engine using 22 weak signals fused through a Bayesian belief tracker. |
| 3 | spark-fleet-revenue-engine | Python | Autonomous revenue engine: OCR + LLM sponsor extraction, Apollo enrichment, Zoho CRM sync, WhatsApp/email outreach. |
| 4 | cognitive-traffic-orchestrator | TypeScript / Python | 7-agent AI orchestration for 8,000+ traffic incidents across 22 Bengaluru corridors. LightGBM + RAG, 11 FastAPI endpoints, 1ms brief latency, React/Vite dashboard. Flipkart Gridlock Hackathon 2.0. |
| 5 | Terra-Pulse | HTML/JS | Frontend environmental dashboard / data project. |
| 6 | gmail-browser-agent | Python | Browser-automation agent for Gmail workflows. |
| 7 | azure-multimodal-ingest-engine | Python | Multimodal ingestion pipeline on Azure. |
| 8 | LLM-Facet-Evaluator | Python | Faceted evaluation framework for LLM outputs. |
| 9 | amex-profitability-ranking- | Python | Kaggle-style credit-card profitability ranking solution. |
| 10 | Flipkart-Grid-Traffic-Demand-Forecasting | Python | Demand forecasting for traffic logistics — Flipkart Grid. |
| 11 | NVIDIA-Nemotron-Model-Reasoning-Challenge-Solution-Kaggle | Jupyter | Kaggle solution for the NVIDIA Nemotron model reasoning challenge. |
| 12 | Blog_Agent | Python | Agentic blog research & writing pipeline. |
| 13 | PPT-Generator-AI-Agent-using-Gemini | Jupyter | Gemini-powered agent that generates presentation decks. |
| 14 | Vision_Transformer_from_scratch | Jupyter | Minimal ViT implementation from scratch for learning. |
Active forks & contributions to the LLM / inference ecosystem. I read source, reproduce issues, and upstream fixes.
| Project | Role | Focus |
|---|---|---|
| vllm-project/vllm | Contributor | Diagnosed FP8 MLA KV-cache miscalculation; fixed scheduler KV-cache leak under pipeline parallelism; validated with 7 unit tests. |
| unslothai/unsloth | Contributor | Optimized Triton kernels, fixed LoRA broadcasting bug, patched dynamic sequence padding for multi-GPU, benchmarked 2× fine-tuning speedups. |
| sgl-project/sglang | Contributor | High-performance serving framework with RadixAttention |
| flashinfer-ai/flashinfer | Contributor | Kernel library for LLM serving |
| huggingface/transformers | Contributor | Model-definition framework |
| ggml-org/llama.cpp | Contributor | Efficient LLM inference in C/C++ |
| openvinotoolkit/openvino | GSoC proposal / study | AI PC inference optimization, NPU/GPU/CPU runtime |
| Lightning-AI/pytorch-lightning | Contributor | Scalable PyTorch training |
| ruxailab/sentiment-analysis-api | Contributor | Audio/video sentiment analysis Flask API |
$ git log --oneline --author=InfoSage05 --since="2026-04-01"
[unsloth] optimize Triton kernels to reduce peak VRAM during grad accumulation
[unsloth] fix LoRA weight initialization broadcasting bug
[unsloth] patch dynamic sequence padding for multi-GPU tensor shape mismatches
[vllm] diagnose FP8 MLA KV-cache miscalculation causing memory overestimation
[vllm] fix scheduler bug leaking KV-cache blocks under pipeline parallelism
[vllm] replace unsafe dict lookup with guarded cleanup + 7 unit testsThese are the areas I have been actively contributing to, as detailed in my résumé.
Research Intern · Prof. Deepu John · Jun 2026 – Present
- Architected a 5-layer agentic pipeline combining PostGIS, LLMs, and LightGBM to predict housing prices across 15,000+ properties.
- Built a Python ETL pipeline with PPR and OpenStreetMap connectors plus automated handling of rejected records.
- Engineered an OpenRouter LLM agent layer with exponential backoff and JSON-mode parsing, achieving 90% accuracy.
- Containerized the full stack (Postgres, Redis, FastAPI, React) via Docker Compose for sub-second map visualizations.
Apr 2026 – Present
- Diagnosed FP8 MLA KV-cache miscalculation causing memory overestimation during long-context inference.
- Fixed root cause across two attention spec classes and resolved a related OOM via lazy workspace memory allocation.
- Fixed a scheduler bug where orphaned requests under pipeline parallelism leaked KV-cache blocks and crashed the engine.
- Replaced unsafe dict lookup with guarded cleanup, validating correctness via 7 targeted unit tests.
May 2026 – Present
- Optimized custom Triton kernels to eliminate redundant tensor allocations, reducing peak VRAM during gradient accumulation.
- Fixed a critical broadcasting bug in LoRA weight initializations across variable scaling architectures.
- Formulated an active patch for dynamic sequence padding, resolving tensor shape mismatches during multi-GPU training.
- Benchmarked throughput across quantization configs, verifying 2× faster fine-tuning speeds via integration tests.
Feb 2025 – Jun 2025
- Built RAG pipelines for a GenAI mental-health assistant using HyDE, GRAG, and LCM+RAG techniques.
[2026] Team Vice-Captain — General Championship Data Analytics 2026 (IIT KGP)
Frammer AI · 4th place among 18 teams · Agentic AI analytics with 105+ KPIs
[2026] Qwen Cloud Global AI Hackathon — QwenSwarm RepoPilot
SGLang RadixAttention · 3,500ms latency reduction · 83.3% cache hit rate
[2026] Flipkart Gridlock Hackathon 2.0 — Cognitive Traffic Orchestrator
7-agent system · 8,000+ incidents · 22 Bengaluru corridors · 1ms briefs
[2026] Convolve 4.0 (Pan IIT) — NyayaAI Multi-Agent Legal Assistant
[2025] Winner — General Championship Data Analytics 2025 (IIT KGP)
Market-Mix Modeling · 1.05M records · 129% ROI · 29% Net Profit
[2025] Gold Medal — Marketing Budget Optimization (GenChamp Technology, IIT KGP)
[2025] Amazon ML Challenge — Multimodal Product Pricing Engine
[2025] Google Cloud Agentic AI Hackathon — Project Kisan
[2025] Kharagpur Data Science Hackathon — Research Paper Evaluation Framework
[ONGOING] CGPA 8.41 / 10 — IIT Kharagpur Dual Degree · Micro Spl. AI & Applications
{
"languages": ["Python", "C/C++", "SQL", "JavaScript", "TypeScript"],
"ml": ["PyTorch", "TensorFlow", "Keras", "Scikit-learn", "HuggingFace", "LightGBM"],
"llm_engineering": ["LangChain", "LangGraph", "SGLang", "vLLM", "Unsloth", "OpenVINO"],
"agents": ["MCP", "CrewAI", "Groq", "Gemini", "OpenRouter"],
"backend": ["FastAPI", "Django", "Flask", "Pydantic", "React"],
"infra": ["Docker", "AWS", "GCP", "Azure", "Git/GitHub"],
"databases": ["PostgreSQL", "Redis", "Qdrant", "ChromaDB", "FAISS"],
"systems": ["Linux", "WSL", "Bash", "Git"]
}- 📧 Email: mr.infosagepoint2020@gmail.com
- 🔗 LinkedIn: in/ayushman-paul-b7796a27b
- 🐦 X: @ayushman_p1001
- 🌐 Portfolio: infosage05.github.io/my-portfolio
- 💻 LeetCode: InfoSage05
$ echo "Thanks for stopping by." > Thanks for stopping by. $ exit