Integrated B.Tech (IT) + MBA student at ABV-IIITM Gwalior (Aug 2023 – May 2028). I engineer at the intersection where machine learning meets high-performance systems: retrieval pipelines with zero tolerance for hallucination, graph neural networks evaluated without temporal leakage, and socket-level network protocols you can watch recover under Wireshark packet drops.

The thesis behind everything I build: A system isn't finished when it runs without errors once. It is finished when I can put a hard number on it (latency percentiles, PR-AUC, false-positive rate at fixed recall) and explain mechanically why that number moved.

SELECTED WORK

[ DuckDB ] [ Polars ] [ PyArrow ] [ Redpanda / Kafka ] [ dbt ] [ Parquet / Iceberg ] [ Exactly-Once ]

A high-throughput financial market stream processing engine and Medallion Lakehouse platform built to handle 80,000+ events/sec. Solves the three primary failure modes of real-time streaming architectures: the Lakehouse Small-File Problem via an autonomous compaction worker, at-least-once duplicate corruption via deterministic event-time watermarking, and consumer worker crashes verified via automated chaos testing.

Battle-tested architectural decisions

• Solving the Small-File Problem (CompactionWorker): Micro-batch streaming writes small 20KB–500KB Parquet files that cause metadata thrashing and disk seek bottlenecks. The autonomous compaction engine gathers fragmented partitions exceeding threshold (min_files = 4), merges Arrow tables in zero-copy chunks, and writes consolidated 128MB Parquet files, accelerating analytical query scans by 5.05x.
• Event-Time Watermarking & Idempotent Deduplication: Maintains a monotonic high-water mark with a 30-second bounded delay window, quarantining out-of-order ticks from disparate exchanges while deduplicating replayed primary keys idempotently.
• Chaos Engineering & Exactly-Once Semantics: Forcefully terminates consumer processes mid-flight (SIGKILL) with injected duplicate bursts and late ticks; replayed broker streams recover state with zero duplicate trades and zero lost executions.
• Medallion Architecture & Kimball Data Marts: Raw immutable Bronze append logs promote to cleaned Silver fact tables with Dead-Letter Queue (DLQ) quarantine, feeding Gold marts (agg_intraday_vwap, fct_execution_slippage, dim_venues) with strict dbt data contract assertions.

Python DuckDB Polars PyArrow Redpanda dbt Streamlit Docker Medallion Architecture

[ LangGraph ] [ Qdrant ] [ RAG ] [ grounding verification ] [ FlashRank ] [ $0 Cloud Stack ]

Financial regulatory filings are a notoriously hostile retrieval target: essential numerical answers live inside dense, multi-column financial statements, comparative questions cross arbitrary quarters, and an ungrounded hallucination is worse than a hard refusal. This is an autonomous, multi-hop retrieval and reasoning engine for SEC 10-K and 10-Q documents, orchestrated as a LangGraph state machine over a Qdrant vector store with local cross-encoder reranking.

Battle-tested architectural decisions

• DEC-005 — Database-level metadata isolation over application filtering: Naive vector search over multi-quarter corpora risks retrieving chunks from the wrong quarter if semantic embeddings are close. Enforced ticker, fiscal_year, and fiscal_period as indexed payload filters evaluated at query time inside Qdrant. This prevents cross-period context contamination by construction rather than post-hoc filtering.
• DEC-009 — Table-preserving extraction (pdfplumber): Generic PDF loaders destroy column alignment when extracting financial tables. Implemented table-aware extraction with Markdown table serialization, tagging chunks as type: table so tabular data retains mathematical structure.
• DEC-006 — Local Cross-Encoder Reranking (FlashRank): Narrative disclosures (risk factors, boilerplate) often match raw query keywords better than numbers. Applied a local ms-marco-TinyBERT-L-2-v2 cross-encoder reranking pass in single-digit milliseconds, lifting financial balance sheets to the top of the context window.
• DEC-007 & DEC-014 — Two-pass synthesis with bounded hallucination recovery: A single prompt that generates and "self-checks" its own output routinely confirms its own errors. Decoupled generation into generate_comparison and an independent verify_grounding node. If grounding checks fail, a bounded retry loop (MAX_GENERATION_RETRIES = 1) triggers regeneration with explicit context before safely falling back to handle_missing_data.
• DEC-015 — Bounded backoff and observability: Wrapped LLM chains in tenacity exponential backoff restricted strictly to transient faults (rate limits, timeouts), with structured per-node latency logging.

Python LangGraph Qdrant Cloud AWS EC2 Docker FastAPI Streamlit FlashRank pdfplumber Groq

[ GraphSAGE ] [ PyTorch Geometric ] [ XGBoost ] [ fraud / AML ] [ Optuna ] [ leakage-free ML ]

Money laundering is fundamentally a relational graph problem: individual transactions look completely legitimate in tabular rows, while the illicit pattern exists in cyclic layering, fan-out distributions, and multi-hop mule account chains. Modeled 5M+ IBM transactions as a dynamic graph and trained an inductive GraphSAGE classifier to flag money-mule laundering typologies that conventional rule engines miss.

Battle-tested architectural decisions

• Linear chain graph vs. explosive all-pairs line graph: Connecting all transaction pairs sharing an account would produce ~14 billion edges from the single largest hub account alone (168,672 transactions), exceeding hardware limits. Engineered a chronological chain graph where each transaction links only to its temporal neighbor per account, keeping edge complexity linear (~1.82 average degree) while preserving layering typology sequences.
• Evaluation trap: average_precision_score vs invalid auc(): Discovered that trapezoidal auc(recall, precision) interpolates linearly between PR points, artificially overestimating step-function scores by up to 3.7x on discrete baselines. Switched strictly to Average Precision scoring.
• Strict chronological split (no lookahead leakage): Velocity and temporal features depend on history. Standard random cross-validation leaks future transactions into historical windows. Enforced strict 70/15/15 chronological quantile splitting (3.5M train / 761K val / 761K test).
• Dataset artifact detection: Identified an uncharacteristic background traffic collapse in the raw IBM dataset after Day 10; restricted training window to Days 1–10 to avoid training on degenerate synthetic artifacts.
• Optuna GNN optimization: Evaluated neighbor aggregation functions (mean, lstm, max-pooling) under a 4GB VRAM constraint (RTX 3050), opting for GraphSAGE over memory-intensive GAT attention layers.

Python PyTorch Geometric GraphSAGE XGBoost Optuna scikit-learn temporal validation PR-AUC

[ C++17 ] [ Winsock ] [ UDP sockets ] [ protocol design ] [ game theory ] [ Wireshark ]

A decentralized peer-to-peer file distribution protocol written in C++17 from the socket layer up, with zero external networking libraries and no centralized tracker. Because raw UDP provides neither ordering nor delivery guarantees, protocol reliability, piece scheduling, anti-leeching game theory, and disk persistence were engineered directly into the software architecture.

Battle-tested architectural decisions

• Decoupled threading model (SocketManager): Dedicated background listener thread runs a blocking recvfrom() loop, deserializing datagrams into Packet structures and pushing to a thread-safe queue. The main loop polls at 10 ms intervals, completely isolating socket IO from scheduling state machines.
• Zero-allocation direct-to-disk persistence (PieceManager): Rather than caching multi-megabyte payloads in memory, initialize_file_layout() pre-allocates zero-filled files on startup. Inbound chunks are written directly to exact offsets via seekp(piece_idx * piece_size) under mutex protection.
• Rarest-first swarm distribution: Peers evaluate connected bitfield rarity counts to request least-frequent pieces first, preserving rare chunks from dropping offline if seeders exit.
• Bandwidth reciprocity (ChokeManager): Evaluates upload vs. download byte ratios per peer; sends CHOKE frames to uncooperative nodes to incentivize balanced contribution.

C++17 Winsock raw UDP datagrams CMake Wireshark GoogleTest Streamlit multithreading

──────────────────────────────  A R S E N A L  ──────────────────────────────

• Algorithmic Problem Solving: 500+ Data Structures and Algorithms problems solved in C++ on LeetCode.
• Leadership & Communication: Managed technical documentation, presentations, and operations for 5+ flagship events during the Annual Technical Festival at ABV-IIITM Gwalior.
• Academic Distinction: Scored 98.6 percentile in JEE Main 2023 nationwide among 1.1M+ candidates.

────────────────────  C O M M I T   H I S T O R Y  ────────────────────

Let's talk

If you are building systems where latency, correctness, or scalability needs to be measured and proven rather than assumed, I'd love to connect.