Projects

Sentinel

A multi-agent LangGraph system that audits codebases in parallel, grounding every finding in real tool output rather than LLM pattern-matching.

  • Security, Quality, and Test analysts built on semgrep and pip-audit output.
  • Fixed an unbounded-recursion bug (50+ tool calls) by adding per-analyst step budgets and real-time cost caps.
  • Rekeyed idempotency dedup on stable code identity instead of non-reproducible LLM free text, verified zero duplicate PRs across repeat live runs.
  • Human-review boundary caught two incorrect auto-opened PRs from an out-of-scope finding before merge.

Roadvision Kaggle Datathon

A two-model YOLOv8 ensemble for 13-class vehicle detection on CCTV footage, placing top 50 on the leaderboard.

  • Trained on 810 images with 10,475 annotations.
  • Designed a temporal, per-camera train/val split to prevent near-duplicate frame leakage.
  • Fused predictions via Weighted Box Fusion at 1280px with test-time augmentation, improving mAP over single-model NMS.

Valerix

Champion, BUET CSE Fest 2026. Order and Inventory backend services built for fault tolerance under load.

  • 2-second timeout on synchronous inter-service calls, falling back to async RabbitMQ messaging with an immediate queued status.
  • Idempotent order processing to prevent double inventory deduction on duplicate messages.
  • Indefinite retry-on-connect logic so both services self-heal after RabbitMQ or startup failures.
  • Instrumented with Prometheus metrics and a fault-injection flag used to validate the timeout/fallback path end-to-end.

grít

A developer friction logger — a CLI that hooks into git and your editor to capture the moments that slow you down, and turns them into a searchable local knowledge base.

  • Built adaptive pre-commit interviews in Go: large diffs get split-commit prompts, new files get contract questions, revert commits trigger a 3-question post-mortem.
  • Built a real-time file watcher (grit watch) that flags vague identifiers, large AI-assisted pastes, and long inactivity, using a language-aware naming heuristic and a keyword-based complexity score.
  • Designed a local-first SQLite store (WAL mode) under .grit/ in each repo, with no cloud dependency.
  • Built grit stats (weekly summaries, per-file complexity trends, a 12-week friction heatmap) and grit push for Markdown/JSON export.
  • Wrote idempotent git hook installation (pre-commit, post-rewrite, post-commit) that no-ops safely if the binary isn't on PATH.

HealthX

An AI-powered, offline-capable telehealth platform built for Bangladesh's hill tracts and rural regions, built for FutureBuilders 2025.

  • Built an AI symptom scanner that combines a patient's current input with their medical history for context-aware diagnosis, using Gemini.
  • Built a hybrid online/offline architecture: live Gemini analysis and OpenStreetMap hospital lookup when connected, falling back to a hardcoded local dataset when the connection drops.
  • Built a smart hospital finder using GPS and the Overpass API, with one-tap Google Maps/Waze navigation to the nearest facility.
  • Compressed images client-side before upload so the app stays usable on unstable 2G/3G connections.
  • Built a digital OPD system for asynchronous consultation requests with human specialists, suited to areas with unreliable internet.

mini-pdf

A command-line PDF compressor written in Go, with a live terminal UI for comparing sequential and concurrent execution.

  • Built dual execution modes (sequential and concurrent) using a worker-pool pattern, run side by side by default to compare speedup.
  • Pre-scans input PDFs and sorts them in descending file size before queuing, so the largest files reach workers first and don't create a late-stage bottleneck.
  • Built a live-updating terminal UI with Bubbletea, Bubbles, and Lipgloss showing progress bars, spinners, and styled logs.
  • Reports storage reduction (MB/KB and percentage) and time saved, with a measured speedup multiplier between modes.
  • Used pdfcpu as the underlying compression engine, wired through a channel-based job queue.

CurioKids

A gamified educational mobile app for young learners, with parent and child dashboards, interactive mini-games, and an AI chatbot.

  • Built a React Native (Expo) app with dual dashboards: analytics and task management for parents, gamified courses and rewards for children.
  • Built interactive mini-games (Balance Scale for math/logic, Bubble Pop for letter recognition, Tracing for fine motor skills).
  • Built a Node.js/Express backend syncing progress and analytics to Firestore via the Firebase Admin SDK.
  • Integrated a Retrieval-Augmented Generation chatbot, served locally through Ollama and Docker Compose, as an interactive learning companion.
  • Built a points, leaderboard, and friends system to drive engagement.

The Existential Choreography Planner

A team project (university coursework) — a gamified task manager where chores become drag-and-drop flowcharts with energy costs and point rewards. I worked mostly on the frontend, with some backend work.

  • Built the React Flow-based flow editor: drag-and-drop task nodes, dependency edges, and a side panel for editing task details.
  • Implemented conditional UI logic — locked nodes for incomplete dependencies, a disabled 'Complete Task' state when energy is insufficient.
  • Worked on parts of the Express/MongoDB backend, including the flows and tasks API endpoints.
  • Built the gamification-facing UI: energy bar, points/level display, and badge gallery.
  • Built JSON import/export for flows, including read-only preview before saving to an account.