B.Tech in Information Technology @ Vishwakarma Institute of Technology (VIT Pune)
Focused on Autonomous Agents, Production RAG Systems, LLMOps, and High-Performance Backends.
I build production-grade AI systems, evidence-grounded retrieval pipelines, and backend architectures. Rather than stitching together fragile API wrappers, I engineer from first principles โ from mathematical loss functions and numerical optimization to deterministic agent state loops and AST sandboxing.
problem โ mathematical foundations โ system architecture โ prototype โ benchmark โ harden โ ship
Recent work is tracked here to keep the profile tied to concrete engineering output rather than static skill claims.
- 2026-09-21 โ LeetCode #1971: Find if Path Exists in Graph: added and compile-tested an iterative DFS solution using an adjacency list, with O(n + e) time and space analysis.
- 2026-09-20 โ LeetCode #102: Binary Tree Level Order Traversal: added breadth-first traversal with queue-based level grouping and documented the O(n) / O(w) complexity.
- 2026-09-19 โ LeetCode #70: Climbing Stairs: added a constant-space dynamic-programming solution and documented the state-compression pattern.
- 2026-09-18 โ Sypher task cancellation regression test: added coverage for pending-task cancellation, dispatch exclusion, and repeated-cancel behavior.
|
Building autonomous decision systems combining multimodal perception (voice + screen OCR), reasoning graphs, dynamic tool registries, and AST-level safety execution gates. |
Engineering enterprise retrieval pipelines with hybrid lexical/dense search (BM25 + Dense), Reciprocal Rank Fusion (RRF), cross-encoder reranking, and empirical evaluation (Recall@K / MRR). |
Building packet analysis engines, stateful TCP flow trackers, and heuristic intrusion detection systems, alongside low-level C++ algorithmic patterns and mathematical ML optimization from scratch. |
Bridging multimodal LLM reasoning directly with real desktop operating system execution.
- Multimodal Perception: Listens via voice, perceives the active display via screen capture + OCR, and interprets visual context in real time.
- AST Security Gate: Intercepts generated code using Python's Abstract Syntax Tree (ast) to eliminate arbitrary injection vulnerabilities before system execution.
- Stateful Execution: Features persistent episodic memory, user approval thresholds for sensitive operations, and dynamic task queuing.
Voice / Screen Perception โ Reasoning Engine โ AST Security Gate โ OS Execution Engine โ Memory Feedback Loop
Evidence-first document intelligence with hybrid retrieval, reranking, citations, abstention, and golden evaluation.
- Hybrid Search Engine: Fuses exact lexical matching (BM25 / TF-IDF) with semantic vector search using Reciprocal Rank Fusion (RRF).
- Cross-Encoder Reranker & Abstention: Re-scores top candidates and enforces a strict confidence threshold โ deliberately abstaining when evidence is insufficient to eliminate hallucinations.
- Deterministic Evaluation: Ships with a golden evaluation harness computing Recall@5 and Mean Reciprocal Rank (MRR), complete with CI automation.
Documents โ Chunker โ [BM25 + Dense Search] โ RRF Fusion โ Cross-Encoder โ Abstention Check โ Grounded Citations
| Repository | Focus Area | Key Architectural Signal | Stack |
|---|---|---|---|
| ๐ง Sypher | Autonomous Agent | Multimodal screen perception, Gemini reasoning, and AST-secured OS execution | Python โข Gemini โข FastAPI โข WebSockets |
| ๐ง Agentic Research System | Autonomous AI Agents | Directed state graph, self-reflection critique loops, AST sandbox & OpenTelemetry tracing | Python โข State Graphs โข Tracing โข CI/CD |
| ๐ Grounded RAG Lab | Enterprise Retrieval | Hybrid BM25/Dense search, RRF fusion, cross-encoder reranker & Recall@5 eval | Python โข BM25 โข RRF โข FastAPI โข CI/CD |
| ๐ Sentinel IDS | Network Security & IDS | Zero-dependency L2-L7 packet decoder, stateful TCP flow tracking, Port scan, SYN flood, ARP spoofing & DNS entropy heuristics | Python โข TCP/IP โข PCAP โข Network Forensics โข CI/CD |
| ๐ ML Foundations | ML From Scratch | Gradient descent, MSE loss, binary cross-entropy, custom standardizer & metrics | Python โข NumPy โข Linear Algebra |
| ๐ Semantic Search Lab | Information Retrieval | Inverted index vocabulary, TF-IDF weighting, L2 normalization & cosine ranking | Python โข NumPy โข FastAPI |
| ๐๏ธ Vision Inspector | Computer Vision | Classical Canny edge detection, contour geometry, circularity metric & CLI | Python โข OpenCV โข Image Processing |
| ๐ง LeetCode | Algorithmic Problem Solving | Pattern-oriented problem catalog (Two Pointers, Hashing, DFS, Heaps) in C++ | C++17/20 โข Data Structures โข Algorithms |
- Core Languages: Python, C++ (17/20), C, Java, SQL, JavaScript/TypeScript
- AI, ML & Retrieval: RAG Architectures, Hybrid Search (BM25 + Dense), Cross-Encoders, PyTorch, scikit-learn, OpenCV, NumPy, Pandas, Transformers
- Networks & Security: TCP/IP (L2-L7), Raw Sockets, PCAP Analysis, Network Forensics, Intrusion Detection, ARP/DNS Security
- Backend & DevOps: FastAPI, RESTful APIs, WebSockets, Docker, Git/GitHub, Linux Environments, GitHub Actions CI/CD
- Systems over Wrappers: An LLM without a deterministic state graph, security boundaries, and evaluation is just a toy. True AI engineering lives in the architecture around the model.
- Empirical Evaluation: If you can't quantitatively measure your retrieval quality (Recall, MRR), latency, and failure cases, you cannot improve them.
- Mastery from First Principles: Understand the math, the loss surface, and the gradient equations before importing high-level library abstractions.

