Focusing on Generative AI, Agent Infrastructure, LLM Inference Engines, and Distributed Storage Systems.

I'm an AI & Systems Engineer who loves working inside complex infrastructure, solving deep engineering problems, and contributing improvements upstream. My technical work spans machine learning frameworks (PyTorch), high-throughput LLM inference engines (vLLM), financial AI workflow engines (DealLens), distributed P2P storage engines (Tejas-DB), and Linux ecosystem packaging toolchains (Canonical Snapcraft & Craft Parts).

• 🎓 Education: B.Tech in Computer Science & Engineering @ NIT Durgapur ('27)
• 🧠 Primary Focus: Generative AI, LLM Inference Optimization, Agent Runtimes, and Distributed Storage
• 🛠 Open Source: Active upstream contributor to vLLM, PyTorch, Canonical (Snapcraft & Craft Parts), CP Editor, and Automattic Jetpack (9 Merged PRs)
• 🏆 Competitive Programming: 4★ on CodeChef (Max Rating: 1804)

Merged PR #49206: Fix request index preemption misalignment in SchedulingPolicy.PRIORITY

• Problem: Under heavy KV cache memory pressure, request preemption in SchedulingPolicy.PRIORITY resulted in index misalignment within waiting request queues, causing lower-priority or preempted requests to be silently skipped during re-scheduling.
• Solution: Re-engineered preemption queue index offset calculations in vllm/core/policy.py, preserving strict priority queue ordering during KV cache page eviction and re-insertion.
• Tech Stack: Python, PyTorch, LLM Inference, KV Cache Eviction

Merged PR #189142: Fix return_annotation schema for tuple-returning operators in PyTorch FX

• Problem: Incorrect schema typing for tuple-returning operators in PyTorch FX graph tracer caused static type checkers and downstream graph transformations to crash.
• Solution: Corrected FX operator schema definitions to accurately return tuple type annotations across symbolic graph execution paths.
• Tech Stack: Python, FX Graph Tracer, Compiler Schemas

Merged PR #190191: Fix floating-point division-by-zero crash in sparse_compressed_to_dense

• Problem: Executing conversions on malformed BSR compressed sparse tensors triggered hard C++ floating-point division-by-zero runtime segmentation faults.
• Solution: Added strict boundary validation and non-zero checks in underlying C++ sparse tensor conversion kernels.
• Tech Stack: C++, Sparse Tensors, Error Handling

Craft Parts — Merged PR #1628: Prevent file deletion during self-linking in link_or_copy

• Problem: In link_or_copy, when source and destination paths resolved to the same physical file (e.g., via staged symlinks), catching EEXIST and unlinking the destination accidentally unlinked and permanently deleted the source file itself (Fixes canonical/snapcraft#6168).
• Solution: Implemented a physical file comparison check (os.path.samefile) in craft_parts/utils/file_utils.py to safely return early without deleting the target asset.
• Tech Stack: Python, Linux File Systems, Symlinks, Packaging Subsystems

Snapcraft — Merged PR #6272 & PR #6269

• Solution: Resolved hardcoded linter fallbacks across package build pipelines via dynamic build_base resolution, and contributed manual connection documentation for the personal-files security interface.
• Tech Stack: Python, Snapcraft CLI, Security Interfaces

Merged PR #1501, PR #1499, & PR #1498

• LLM Localization Pipeline (PR #1501): Translated 3,600+ UI strings using an automated LLM translation pipeline and fixed duplicate search indexing in global search panels.
• Dynamic Font Scaling & Tab Controls (PR #1499, #1498): Added Ctrl+Scroll font scaling toggle and implemented Qt mouse event filters for middle-click tab closure.
• Tech Stack: C++17, Qt Framework, LLM Pipeline, GUI Architecture

• Contributed 9 merged PRs fixing Gutenberg block editor state transformation crashes (PR #50035, PR #50025), resolving OpenAPI parser failures via polymorphic oneOf schemas (PR #50030), and refactoring string store IDs to type-safe store objects in React hooks (PR #49810).

+-----------------------------------------------------------------------------------------------+
|  Tejas-DB (Distributed Key-Value Database in C++17)                                           |
|  - Consistent Hashing | Gossip Protocol | Tunable Quorum (N, W, R) | WAL Engine                 |
|  - Benchmark: 33,685 req/s throughput | ~2.97 ms avg latency | 100 concurrent threads          |
+-----------------------------------------------------------------------------------------------+
|  DealLens (AI Investment Due-Diligence & RAG Workflow Engine)                                 |
|  - Deterministic 7-step DAG State Machine | PostgreSQL 16 + pgvector Hybrid Search (RRF)      |
|  - Page-aware 1-indexed PDF chunking | Custom CitationVerifier Guardrail                       |
+-----------------------------------------------------------------------------------------------+
|  NexusMatch / AI Job Platform (AI Career Intelligence)                                        |
|  - Decoupled FastAPI/Laravel Backend | scikit-learn TF-IDF & Cosine Similarity Matching        |
|  - Real-time job discovery via async Celery workers + Redis | Groq / LangChain Active Fallback|
+-----------------------------------------------------------------------------------------------+

Decentralized P2P Storage Engine built in C++17

• Architecture: Engineered a decentralized peer-to-peer key-value storage system utilizing Consistent Hashing for dynamic data partitioning and a Gossip Protocol for autonomous cluster node discovery and failure detection.
• Concurrency & Reliability: Implemented a concurrent engine using std::shared_mutex, customizable Quorum Consensus (N, W, R) for tunable consistency, and a Write-Ahead Log (WAL) engine for instant crash recovery.
• Performance Benchmarks: Achieved 33,685 req/sec throughput at ~2.97 ms average latency under 100 concurrent thread workloads. Includes dynamic Vercel Native Edge Proxy rewrites for active backend tunnels.

AI Investment Due-Diligence & Financial RAG Engine

• Workflow State Machine: Designed an asynchronous DAG workflow engine (Validation → Entity Extraction → Performance Analysis → Risk Analysis → Evidence Retrieval → Claim Verification → Report Generation) replacing non-deterministic agent loops.
• Hybrid RAG & Provenance: Built a PostgreSQL 16 + pgvector hybrid search engine combining dense Cosine Distance embeddings and sparse tsvector keyword search via Reciprocal Rank Fusion (RRF). Integrated a custom CitationVerifier guardrail to enforce strict source provenance and prevent hallucinations in corporate document analysis.

AI Career Intelligence & Resume Matching Platform

• Semantic Matching Engine: Built a high-performance career intelligence platform featuring an optimized TF-IDF and Cosine Similarity engine that evaluates resume compatibility in milliseconds.
• Asynchronous Scalability: Utilized Celery and Redis for asynchronous background tasks, paired with active model fallback loops via LangChain and the Groq API.