About Me  •  Tech Stack  •  Projects  •  Open Source  •  Current Work  •  Activity  •  Connect

I'm an AI & ML undergraduate at Amrita Vishwa Vidyapeetham, Coimbatore.

I love building machine learning systems that are engineered like real products (clean typed APIs, comprehensive test suites, reproducible one-command deploys) and evaluated like serious research (leakage-free data splits, honest ablation studies, statistical significance tests, and a transparent account of what didn't work).

Most of my day-to-day focus sits right where applied ML meets practical systems engineering:

• Critical-infrastructure security: Causal intrusion detection in power grids and industrial SCADA networks.
• Agentic-AI governance: Enforcing deterministic security boundaries (via OPA / Rego) around autonomous LLM agents.
• Payments & financial reconciliation: High-throughput ledger matching with microsecond latencies and zero margin of error.
• Compiler internals: Digging into PyTorch Dynamo, torch.compile, and operator lowering.

💡 Guiding Invariant: "If a model loses to its own simple baseline, my READMEs say so right on the front page alongside the headline number."

Languages & Scripting

Machine Learning & Causal AI

Backend, Systems & Protocols

Security & Simulation

Policy-bounded agentic governance for distributed cloud pipelines.

Autonomous AI agents can fix broken data pipelines on the fly, but letting an LLM execute arbitrary shell or SQL commands in production is reckless. I built ACDE to enforce safety through architecture rather than prompt alignment: four specialized agents continuously monitor pipeline telemetry and propose remediation steps, but an Open Policy Agent (OPA) gate evaluates every proposed mutation against deterministic security policies before anything runs. Replicated and extended an arXiv research paper with a dedicated adversarial containment evaluation suite.  •  556 unit tests · Adversarial containment measured at 1.0 (zero policy leaks)

Three-way payment settlement reconciliation, tied out to the exact paisa.

Built for the Razorpay AI Buildathon 2026. Kosh reconciles high-volume financial transactions across customer orders, payment gateway ledgers, and bank statements in ~40 ms for 10,000 records. It follows a strict deterministic-first philosophy: fast vector math and heuristic rules settle 99.5%+ of entries instantly, while a tool-using LLM agent is only ever invoked for the messy <0.5% edge cases that need contextual reasoning. Designed with zero external runtime dependencies.  •  257 tests · 0.00% false-match rate across 11 random seeds & 7 adversarial test cases

Causal intrusion detection for smart power grids: learned, not hand-tuned; closed-loop, not open.

Standard intrusion detection systems in smart grids treat telemetry like arbitrary time-series data, completely ignoring underlying physical grid equations. In this project, a Graph Neural Network (GNN) perception layer feeds into a Dynamic Bayesian Network, with pandapower simulating real physical electrical loop feedback. To rigorously stress-test its resilience, I trained a reinforcement learning (PPO) adversarial attacker specifically tasked with dodging the detector. Extends an IEEE Access publication with 12 pre-registered experiments, keeping and reporting all negative results.  •  530 tests · 12 pre-registered experiments · Full adversarial RL evaluation

MCP-native API attack-surface & BOLA vulnerability detection engine.

Detects Broken Object Level Authorization (BOLA), hidden endpoints, and credential enumeration by analyzing messy raw API access logs into concrete, citable security evidence. Rather than hallucinating findings, it executes seven deterministic detection heuristics exposed as a full Model Context Protocol (MCP) server. Validated against a real 50,000-line third-party production log corpus.  •  111 tests · 11 MCP tools · 6 resources · 4 prompts

Explainable precision-farming platform: 13 decision modules behind a unified service.

Most machine learning tools in agriculture act as black boxes that growers cannot verify. AgriIntelligence packages 13 agronomic models behind a clean FastAPI service, attaching SHAP or Grad-CAM explanations to every single recommendation so users understand the why. Features strict leakage-free temporal splits, Diebold-Mariano statistical significance tests, and an automated script that compiles an IEEE-style paper directly from trained checkpoints.  •  68 tests · Deployed with Docker on Hugging Face Spaces

Multi-layer security stack for IEC 60870-5-104 industrial SCADA networks.

Legacy industrial SCADA systems often transmit sensitive grid control messages in unencrypted plaintext. I engineered a defense stack combining time-keyed AES-256 encryption with automated 60-second key rotation, paired with an XGBoost intrusion detector trained with SMOTE over 411,000 traffic packets. Validated against active cryptographic attack vectors (CCA, CPA, COA, KPA) on the BUT-IEC104-I benchmark dataset.

I'm an active open-source contributor and enjoy fixing performance bottlenecks and security vulnerabilities in the tools I use every day. Here are some key pull requests I've had merged upstream:

Pull Request	Repository	What Was Solved & Why It Matters
#132860	grafana/grafana	[Core Performance] Diagnosed and eliminated catastrophic regular expression backtracking (\B alternation) in Loki log key parsing, achieving a 5,000x speedup on adversarial inputs.
#3732	apache/dubbo-go	[Concurrency & Data Race] Closed a high-concurrency race condition on router cache using atomic.Pointer under concurrent route generation.
#2200	NVIDIA/garak	[Core LLM Red-Teaming] Corrected prompt setters in paraphrase buffs to maintain conversation history and preserve Conversation type integrity during vulnerability scans.
#4892	huggingface/huggingface_hub	[Security Hardening] Hardened repo_id validation by restricting character patterns strictly to ASCII word characters, eliminating subtle naming ambiguities.
#4884	huggingface/huggingface_hub	[Security Hardening] Prevented directory traversal attacks by explicitly rejecting embedded .. path segments in repository files.
#6858	optuna/optuna	[Docs & Quality] Removed obsolete ChainerMNStudy references and corrected broken Sphinx roles in the official documentation.
#6850	optuna/optuna	[Docs & Quality] Fixed outdated trial failure and NaN reporting examples in the FAQ to prevent user confusion.

• Production ACDE Hardening: Developing a multi-tenant isolation layer and policy sandbox for enterprise agentic pipeline governance.
• Causal vs. Deep-IDS Under Attack: Researching whether causal Bayesian representations retain higher adversarial robustness than deep neural detectors under an adaptive RL adversary.
• klaim: Building an evidence-first algorithmic tool to help quick-commerce vendors contest and reverse unjustified marketplace deductions.
• PyTorch Compiler Internals: Reading through the torch.compile / Dynamo codebase, studying SliceVariable handling, guard generation, and operator lowering.

_{Full contribution graph and commit history are displayed on the profile page above this terminal.}

Whether you'd like to discuss agentic AI safety, causal ML in critical infrastructure, open-source work, or explore potential research/internship roles, feel free to reach out!