BTech CSE (Data Science) @ C.V. Raman Global University · 2024–2028
I build production AI systems at the layer where research papers meet real infrastructure —
GPU kernels, mechanistic interpretability, agentic LLM pipelines, and streaming ML.

Socials

[ Languages ]	[ AI / ML / LLMs ]
[ Infrastructure & Data ]	[ Web & DevOps ]

Agentic interpretability platform capturing residual streams, attention patterns, and SAE features at every step of a multi-turn agent trajectory. Cross-step causal activation patching and representation steering on Gemma-2-2b-it with 16k-width GemmaScope SAEs. Pearson r ≈ 0.91 across IOI circuit mapping. Circuit faithfulness score 0.762.

5-node LangGraph pipeline — Search → Read → Reflect → Cross-Check → Write — with self-correcting reflection loops that eliminate hallucinations before output. Every claim is cited. Contradiction detection built in. Delivers full research reports in under 5 minutes. Sole recipient of the AWS Blog Prize globally — only team to receive it.

OpenEnv-compliant RL environment for training autonomous AI triage agents. 5 difficulty tiers, deterministic reward shaping, stateful episode management, and a DQN baseline agent. Built to spec — ranked Top 800 of 31,000+ global teams. Top 2.5% worldwide.

Custom GPU compute kernels written in OpenAI Triton — FlashAttention-2, RMSNorm, SwiGLU. Fuses memory-bound operations to eliminate HBM read/write overhead by keeping intermediate states in SRAM. Bypasses PyTorch execution overhead with zero abstraction layers. Compatible with Triton Inference Server patterns.

End-to-end QLoRA fine-tuning pipeline on Llama 3.3 8B via Unsloth — 2× faster training, half the VRAM. +34% HumanEval improvement over the base model. Deployed via vLLM on Modal serverless GPU. Exposed as a custom MCP server callable inside Cursor, VS Code, and Claude Desktop — zero marginal API cost per completion.

Production FinTech observability platform. Streaming fraud detection at sub-150ms via Kafka microservices. Tree SHAP per-prediction explainability. SEBI/NPCI compliance RAG with cosine similarity reranking. Evidently AI drift detection with automated Airflow retraining loops. AUC-ROC 0.914 on imbalanced transaction data.

Hybrid LSTM Autoencoder + XGBoost ensemble on 18 signal-level features from TEXBAT datasets. 97%+ recall at under 2% false alarm rate and under 15ms inference. Real-time SHAP attribution identifies which signal anomalies triggered each spoofing alert per-prediction.

Secure agent runtime with recursive AST shell exploit validation — intercepts subshell breakouts before execution. Suffix-delta prompt caching middleware cuts LLM API token costs by 50%. Docker-sandboxed execution for all code paths. Deterministic mock replay for regression testing across 100-payload benchmark suite.

Production-grade collaborative code editor with Y.js CRDT for conflict-free real-time sync across concurrent users. POSIX-sandboxed code execution, dual JWT auth (15-min access + refresh rotation), circuit breaker pattern, range-partitioned PostgreSQL, and Redis Pub/Sub for horizontal scaling.

Distributed movie recommendation and watch-party platform across 8 integrated systems. WebSocket-synced real-time playback, SVD collaborative filtering, BERT personalization, Redis Pub/Sub, Kafka event streaming, JWT-secured admin console. 5,000+ movie TF-IDF recommendation engine with sub-50ms WebSocket latency.