🎓 Final-year B.Tech CSE student at Indian Institute of Information Technology, Kottayam.
🎯 Interested in AI/ML engineering, research, data science, and systems-focused software engineering.
🔧 Open-source contributor to PyTorch, Microsoft Agent Framework, and Google ADK Go.

Multi-Agent LLM Orchestration Platform

A FastAPI orchestration engine that classifies and splits queries into 9 domains, dispatches sub-queries in parallel, applies severity-gated (0–3) judgment with resplit-and-retry, and synthesizes answers through LiteLLM fallback chains. A query rewriter compresses verbose prompts (validated at 0.80 cosine similarity) for a 16.92% average token reduction on LMSYS-Chat-1M, and conversation memory is summarized at 60% of the context window. Subagents cover DSA (CoT output), web search (Tavily), and a 3-judge EvaluatorSubagent for response scoring.

Repository · Live Demo

Hybrid Multimodal Product Search Engine

A search engine for large product catalogs that cleans 147K raw Amazon Berkeley Objects (ABO) listings into 56K indexed products and fuses LoRA-fine-tuned SigLIP 2 dense embeddings (768-dim, trained with PEFT) with SPLADE sparse vectors via Reciprocal Rank Fusion (RRF). Vectors are indexed in Qdrant Cloud, product images and metadata are offloaded to AWS S3, and a FastAPI backend serves queries serverlessly on Modal.

Repository · Live Demo

Volatility Forecasting in Global Assets (TimeMixer)

Engineered a TimeMixer architecture forecasting multi-scale volatility across 4 asset classes (Equities, ETFs, Crypto, Forex) and 5 temporal horizons, computing Yang-Zhang estimators directly from live OHLCV data sourced from Yahoo Finance. Benchmarked against traditional GARCH(1,1) models, achieving a 35% reduction in MAE and 39% improvement in sMAPE across all horizons (12, 96, 192, 336, 720 days), and deployed as a live service via a FastAPI backend.

Repository · Live Demo

• #173989 — Mixed-Precision NaN Propagation Guard · Issue #173885

  FP16/BF16 training causes numbers to overflow to infinity, where (∞ − ∞) subtractions produce NaNs that silently corrupt model outputs. Root-caused this bug in PyTorch (torch.compile) and shipped a CPU/GPU numerical guard that forces the difference to zero instead of NaN, preventing catastrophic model corruption during large-scale training.

  Independently adopted by Amazon's AWS PyTorch team into their production infrastructure (commit a5d7261).

• #185355 — Pipeline Parallelism Hook Accumulation Fix · Issue #185331

  torch.distributed.pipelining reused memory buffers during LLM training, causing hooks to accumulate, leak memory, and corrupt gradients. Root-caused and fixed via isolated independent tensors, restoring stable parallel execution.

• #7772 — Agent Tool-Loop max_duration_seconds Wall-Clock Bound · Issue #7587

  Added a configurable max_duration_seconds bound to the tool-calling loop, previously capped only by iteration and call count. Implemented graceful degradation on timeout, with duration tracked cumulatively across human-approval round-trips.

• #6052 — AWS Bedrock Structured Output Enablement · Issue #5966

  Enabled native AWS Bedrock structured outputs within MAF by implementing the schema translation layer. Devised a precise wire-format resolving community serialization bugs, delivering recursive schema enforcement and multi-shape input handling.

• #7283 — FoundryAgent OPENAI_CHAT_MODEL Inheritance Fix · Issue #7272

  Debugged FoundryAgent's inherited OPENAI_CHAT_MODEL leak causing valid agent-reference requests to fail against the Foundry backend. Developed a two-layer fix combining source-level isolation with defense-in-depth stripping.

• #5947 — MCP Progressive Tool Discovery by #6233 · Issue #5821

  Implemented as_progressive_tools() on MCPTool, enabling progressive tool discovery to reduce context token overhead for large MCP servers exposing 50–200+ tools. Superseded by a MAF member (#6233) for broader implementation than MCP.

• #1729 — Vertex AI Session Memory: Code Execution and File Data Serialization · Issue #1728

  adk-go silently dropped ExecutableCode, CodeExecutionResult, and FileData when converting to the Vertex AI Protobuf format, so code-execution agents lost the code they ran and hallucinated on the next turn. Root-caused the loss to two mapping functions in session/vertexai and implemented the missing round-trip serialization, backed by table-driven tests refactored to cmp.Diff at a maintainer's request.