Computer Engineering student building AI-powered developer tooling — static analysis, code intelligence infrastructure, and MCP-based AI agent orchestration.

I enjoy working inside large, high-bar codebases and building systems that let AI agents reason over and act on real software. My recent work spans code intelligence, agentic AI infrastructure, applied ML, and database engineering — with a merged fix into PyTorch's GPU kernel layer and active review-cycle contributions to a 33K-star code-intelligence engine.

• Building AI agent tooling on the Model Context Protocol (MCP)
• Static analysis, AST parsing, and call-graph resolution
• Multi-agent system design for self-healing, resilient infrastructure
• Contributing to production-grade open-source codebases

An MCP-based developer tooling platform that lets AI agents (CLI, IDE, or remote) reason over a codebase and turn a raw problem statement into phased implementation plans and traceable engineering artifacts.

• Built dedicated research, planning, and workflow subsystems with human review gates at every phase transition.
• Designed a versioned contract and adapter layer (request/response envelopes, capability-negotiation manifest) so 6 client surfaces — CLI, IDE, MCP, AI, REST, and Desktop — share a single core engine across Gemini, Anthropic, and OpenAI-compatible providers without breaking changes.
• Shipped 17 sequential development phases (repository init through Release Candidate Stabilization); ran a full-repo audit that surfaced and fixed 27 issues across 7 dependency-ordered batches, including crash-safe atomic writes and AI-provider adapter fixes.

Stack: Python • Pydantic • MCP • Multi-Provider AI Orchestration Links: GitHub

A self-healing traffic digital twin that autonomously detects sensor failures and reconstructs missing spatio-temporal traffic data in real time.

• Designed a multi-agent pipeline (Health Monitoring, Reconstruction, and Digital Twin Layer agents) for continuous, resilient traffic-network observability.
• Benchmarked a spatio-temporal LightGBM model on the real-world METR-LA dataset, achieving 2.48 mph MAE / 6.06% MAPE — a 78.55% improvement over fallback methods — while sustaining 97% observability under 40% simulated sensor failures.

Stack: Python • TypeScript • LightGBM • Graph Neural Networks • Multi-Agent Systems Links: GitHub • Live Demo

An NLP platform that analyzes Hacker News discussions to identify emerging technology trends.

• Clustered 2,000+ Hacker News discussions into 85 semantic topics using Sentence Transformers, UMAP, and HDBSCAN.
• Built forecasting pipelines using Linear Regression, XGBoost, and LSTM with MLflow experiment tracking, improving forecast reliability by 18%.
• Generated natural-language trend summaries using a Gemini-powered prompt pipeline.

Stack: Python • PyTorch • Sentence Transformers • HDBSCAN • UMAP • MLflow • Streamlit Links: GitHub • Live Demo

An interactive, browser-based 3D digital twin and operations control dashboard for a lunar Electrodynamic Dust Shield (EDS) mission concept, simulating orbit navigation, surface landing, and real-time dust-clearing physics on solar panels — paired with production-ready hardware to build the real thing.

• Built a Three.js simulation of AC traveling-wave electrostatic clearing physics, using raycasting and tangential-plane projection so dust particles slide and detach realistically across sloped panel surfaces without artificial bounding boxes.
• Modeled net energy recovery with a Beer-Lambert exponential attenuation model relating dust coverage to solar transmission, driving real-time efficiency and power-recovery telemetry in the dashboard.
• Designed and laid out two physical PCBs (a 12V→1-5kV HV flyback driver and a 3-phase electrode sequencer with 4kV-isolated optoisolation) and wrote ESP32-C3 firmware to drive the traveling-wave electrode sequence end-to-end.

Stack: Three.js, JavaScript (ES6+), C++/Arduino (ESP32-C3), EasyEDA Links: GitHub • Live Demo

Production-oriented database monitoring framework developed during my internship at DCM Infotech.

• Built a modular health monitoring engine using SQL Server DMVs and Extended Events to analyze index fragmentation, storage utilization, transaction log health, and deadlock telemetry.
• Designed a configuration-driven architecture with reusable T-SQL views, stored procedures, historical snapshot collection, and executive health reporting.
• Developed a reproducible validation laboratory covering fragmentation, storage, Write-Ahead Logging (WAL), recovery models, deadlock detection, and end-to-end dashboard validation.

Stack: SQL Server • T-SQL • DMVs • Extended Events • Database Administration

I enjoy contributing to production-grade open-source software and learning from large engineering codebases.

codebase-memory-mcp (33K+ Stars) — Submitted three fixes to the C-based call-graph resolution engine: a Java enum-method extraction fix (#984, maintainer-praised as "genuinely good work," pending final review) and two receiver-qualified CALLS-edge precision fixes targeting false-positive resolution on unrelated method names (#893, #1128), currently iterating with the maintainer on test coverage.

PyTorch (PR #190481 — Merged) — Diagnosed and fixed a correctness bug where torch.nextafter silently returned its input unchanged for bfloat16 on Apple's MPS backend; implemented a bitwise stepping kernel fix merged into the PyTorch main branch.

AST/tree-sitter • Call Graph Resolution • C • C++

Model Context Protocol (MCP) • Multi-Provider AI Orchestration • Gemini API • Claude API

Python • C/C++ • JavaScript • TypeScript • SQL

PyTorch • LightGBM • Graph Neural Networks (GNNs) • Scikit-learn • XGBoost • Sentence Transformers • HDBSCAN • UMAP

FastAPI • Next.js • PostgreSQL • PL/pgSQL • Docker • Pandas • NumPy

SQL Server • PostgreSQL • Oracle Database • T-SQL • PL/pgSQL • Database Administration • Query Optimization • Backup & Recovery • Performance Monitoring

MLflow • Git • GitHub Actions (CI/CD) • Hugging Face • Firebase • Vercel

• 🏆 2nd Runner-Up — Bharatiya Antariksh Hackathon (ISRO), NRSC Hyderabad 2026 — Top 30 of 15,000+ registered teams
• 🏆 Winner — International AI For Sustainability 2026 (1st Place among 200+ international teams)
• 🏆 Top 100 — Google AI Agent Builder, Google Office Bengaluru 2026

• LinkedIn: https://linkedin.com/in/sahil-manglaa
• Email: sahilmangla.official@gmail.com