BK
loading workspace

AI & Backend Engineer · Dhaka, Bangladesh

Bishwajit Kumar Chakraborty

I build multi-agent LLM systems

Software engineer working where agentic AI meets production backends — LangGraph orchestration, FastMCP servers, and RAG pipelines that hold up under real clinical data. Currently shipping secure agentic workflows at Cloudly Infotech Limited, and author of a full-stack GitHub MCP Suite.

View selected work
  • 0 MCP tools shipped
  • 0 tests across 15 modules
  • 0 CGPA · CSE final year
langgraph · orchestrator.py

$ uvicorn app.main:app --reload

INFO FastAPI up · 0.0.0.0:8000

INFO graph compiled → router ▸ 4 nodes

TRACE query="persistent cough, 3 weeks"

ROUTE router → diagnostic_agent

RAG pgvector hybrid · top_k=8 · 41ms

RAG breadcrumb chunks matched: 6/8

ROUTE diagnostic → treatment_agent

FHIR patient bundle resolved · OpenEMR

DONE streamed 1.2k tokens · 0 retries

LangGraph
FastMCP
pgvector

01 About

I like the unglamorous half of AI: retrieval quality, routing bugs, migrations.

Portrait of Bishwajit Kumar Chakraborty
Open to AI / Backend roles

I'm a final-year Computer Science student at Daffodil International University and an AI & Backend engineer at Cloudly Infotech, where I work on a clinical decision-support platform built from LangGraph agents, RAG, and FHIR/OpenEMR data.

Most of my wins come from chasing the boring details — a router misclassifying a query, a PDF whose headings only exist as font sizes, a vector search that needed lexical help. I care about systems that are testable, observable and secure: 80+ tests behind CI gates, encrypted token storage, and RBAC by default.

  • Agentic systemsMulti-agent graphs, tool calling, streaming responses
  • Retrieval engineeringSemantic chunking, hybrid vector + full-text search
  • Production backendsFastAPI, async Python, Postgres, Docker, CI
  • Ships end-to-endFlutter & React clients on top of my own APIs

02 Experience

Currently in production, on a clinical AI platform.

Feb 2026 — Present Full-time · On-site · Dhaka

Software Engineer Intern — AI & Backend

Cloudly Infotech Limited

Clinical decision-support platform · LangGraph agents · RAG · FHIR / OpenEMR

  • Engineering a production multi-agent LLM platform on LangGraph and FastAPI; fixed orchestrator defects that misrouted queries between the Diagnostic and Treatment agents (system prompts, conditional routing nodes, core enums).
  • Led a RAG ingestion overhaul — heading-hierarchy breadcrumb chunking, table-aware atomic chunking, list-boundary splitting, and font-size heading inference for unstructured PDFs.
  • Added hybrid vector + full-text search over pgvector (Postgres tsv generated column, Alembic migrations incl. a multi-head merge); built a MinIO/S3 ingestion pipeline with an async aioboto3 module and a resumable, idempotent uploader CLI.
  • Improved FHIR-backed patient search (exact full-name, title-inclusive matching) and clinical-note PDF export; authored 80+ tests across 15 modules under ruff/black CI gates.
LangGraphFastAPIpgvectorAlembic aioboto3FHIRpytest
2022 — 2026 Daffodil International University

B.Sc. in Computer Science & Engineering

CGPA 3.72 · all coursework completed

Alongside coursework: LoRA/QLoRA fine-tuning experiments, an IoT + AI water-quality system presented to industry leadership, and an NLP/ML research-lab workshop.

03 Selected work

Eight systems I built end to end.

01

FastMCP · FastAPI · pgvector · Flutter

GitHub MCP Suite

Full-stack GitHub automation with AI

  • Production-grade MCP server — 12 tools over FastMCP/SSE for repository management, GitHub Projects v2 (GraphQL), branch/PR creation, issue tracking, and RAG-powered codebase Q&A using pgvector for semantic code retrieval.
  • BFF API layer (FastAPI + SQLAlchemy + Alembic + PostgreSQL) with Firebase Auth, GitHub OAuth 2.0, JWT signing and Fernet-encrypted token storage for secure multi-user credential management.
  • Cross-platform Flutter app with Riverpod state management — MCP tools and GitHub Projects from one interface, with real-time streamed agent responses.
MCPSSEGraphQLFirebase Auth FernetRiverpod
Repository

02

Python · PostgreSQL · FastAPI · FHIR R4 · Docker

MIMIC-IV Clinical Data Platform

De-identified ICU records → FHIR R4 bundles → clinical dashboards

  • Loads MIMIC-IV v3.1 and MIMIC-IV-Note v2.2 (41 tables across hosp / icu / ed / note, 888M rows) into PostgreSQL and converts them to FHIR R4 transaction bundles — 27 resource mappings spanning Patient, Encounter, Condition, MedicationRequest / Dispense / Administration, Procedure, Observation, DiagnosticReport and DocumentReference.
  • Built a second “GoldData” pipeline that blinds only the most recent hospital admission and ED stay — diagnoses, medications, procedures and notes stripped from the latest visit while every prior encounter stays intact, yielding evaluation sets for clinical decision-support models.
  • Tuned for scale: composite indexes plus per-patient caps (500 labs, 2,000 chart events) take single-patient bundle generation from minutes to milliseconds on a 432M-row chartevents table.
  • Ships as a Dockerised PostgreSQL 16 stack, a subcommand CLI (load / reindex / bundle / golddata) and two FastAPI dashboards sharing a five-tab clinical UI — overview, chief complaint, diagnosis & treatment, encounter timeline and searchable notes.
FHIR R4MIMIC-IVETLPostgreSQL 16 NDJSONFastAPIDocker Compose
Repository

03

Express 5 · Mongoose 9 · JWT · Zod · Docker

Secure Voting System

Gono Vote System — election backend with RBAC

  • 11-endpoint REST API covering the full election lifecycle — voter signup and login, profile and password management, admin CRUD over candidates, and public vote tallies. JWT (HS256, 6h) auth via a Passport local strategy, with role-based access control splitting voter and admin routes at the middleware layer.
  • One vote per user, enforced atomically with rollback if the candidate lookup fails; bcrypt at 10 salt rounds with passwords held select: false and compared in constant time; unique voter-ID constraints and Zod schema validation on every payload.
  • Winston logging across four transports — combined, error, uncaught exceptions and unhandled rejections — plus Morgan request logs, graceful MongoDB shutdown, and a Dockerfile / docker-compose deployment with its own test suite.
JWT · HS256RBACPassportbcrypt ZodWinston
Repository

04

Node.js · MongoDB · Flutter · Gemini API

Water Quality Monitoring

LoRa sensor network + AI analysis

  • End-to-end IoT pipeline collecting pH, turbidity, TDS and flow rate from LoRa sensors; Express backend stores schema-validated time-series readings in MongoDB with automatic threshold alerting.
  • Integrated Gemini API for anomaly detection over multi-parameter sensor data, generating human-readable health assessments.
  • Flutter real-time dashboard — live feeds, historical trends and push alerts; presented to industry leadership at DIU.
IoTLoRaTime-seriesGemini

05

Python · Ollama · Tkinter · psutil

Local LLM Desktop Assistant

Offline AI assistant with a live system monitor

  • Fully local assistant on Ollama with zero cloud dependency; a modular plugin architecture hot-swaps LLM backends and capability extensions without restarting the app.
  • Real-time CPU / RAM / GPU telemetry overlay via psutil and platform-native APIs — metrics stream to the UI and can be injected into the model's context for hardware-aware answers.
  • End-to-end voice interaction through STT/TTS pipelines with configurable wake-word detection, runtime model switching and conversation history.
Local LLMPluginsSTT / TTSTelemetry
Repository

06

Flutter · Dart · Provider · GraphQL

AnimeVerse

Cross-platform anime discovery & tracking app

  • Flutter app across six targets (Android, iOS, web, Linux, macOS, Windows) — 12 screens behind a five-tab shell, built on a layered models / services / providers / widgets architecture with Provider + ChangeNotifier state split across auth, anime, favorites and theme.
  • Dual-API data layer — Jikan REST drives trending, genres, search and detail views, with automatic AniList GraphQL fallback when Jikan returns nothing, MAL → AniList ID resolution, and optional watchlist writes behind a per-user access token.
  • Built to degrade gracefully: local auth with session restore, per-user favorites with JSON import/export, recent searches and Material 3 theme persisted via SharedPreferences/Hive, posters that retry across candidate URLs before a placeholder, and explicit loading / error / empty states throughout.
FlutterProviderMaterial 3 GraphQLHive
Repository

07

C99 · OpenGL 2.x · GLUT · OpenAL

Ninja Shadow

2D platformer engine written from scratch in C — no game engine, no textures

  • Every character, enemy, trap and background element is drawn from raw OpenGL primitives — quads, triangle fans, line strips — as black silhouettes. Ten levels load from a text-based level format (platform types, enemy patrol bounds, traps, collectibles, checkpoints) across five Japanese-themed environments with a four-layer parallax backdrop.
  • Hand-rolled physics and collision: gravity integration with clamped terminal velocity, AABB overlap tests with push-out resolution, one-way / moving / crumbling platforms that carry the player, and a capped timestep so fast bodies can't tunnel at low frame rates.
  • Two state machines drive the action — nine player animation states derived from physics (idle, run, jump, fall, wall-slide, attack, dash, hurt, dead) and a patrol / chase / attack FSM across three enemy archetypes — alongside a 256-slot particle pool and an eased tracking camera.
  • OpenAL audio that degrades gracefully: 22 sound effects and 5 music tracks over a 16-source pool, silently no-oping when OpenAL is absent so the game still runs. Built with a GNU Makefile and covered by assert.h unit tests on projectile count, velocity, range and key-repeat suppression.
C99OpenGLGLUTOpenAL AABB physicsState machinesMakefile
Repository

08

Python · Tkinter · psutil · cryptography · pyudev

Linux Security Dashboard

Seven defensive security tools behind one desktop console

  • File integrity monitoring over user-selected paths using SHA-256 checksums, with timestamped tamper logging and alerts the moment a watched file changes.
  • Live system and network telemetry — CPU, RAM and throughput (KB/s in / out) charted on a one-second refresh via psutil and matplotlib, beside a table of active connections showing protocol, local and remote address, state and owning PID.
  • Privileged operations, wrapped safely: iptables/ipset rule management including domain blocking that resolves every A record before adding rules, Fernet symmetric encryption with key generation, and pyudev-driven detection of USB plug/unplug events enriched with udevadm device metadata.
  • Rounded out with a password auditor — length, case, digit and symbol scoring plus a strong-password generator with clipboard handoff.
SHA-256Fernetiptables / ipset pyudevpsutilmatplotlib
Repository

Also built

Crypto Tracker — live market dashboard

Top-coin dashboard on the CoinGecko API — instant search, five sort keys, grid/list toggle, 10-second auto-refresh, and a 7-day price chart drawn by hand with the Canvas 2D API, no charting library, redrawn on resize.

Vanilla JSES ModulesCanvas 2DCORS fallback

RanPass — random password generator

Configurable length (6–16), optional numbers and symbols, click-to-copy with tooltip confirmation, and a peppering reminder. Deployed to GitHub Pages from an Actions workflow.

Tailwind CSSVanilla JSGitHub Actions

OS & systems utilities

Eight standalone Linux tools in Bash and Python: DHCP server provisioning with a health-check probe, a proxy server, a taskmaster process manager, cross-platform CPU/RAM/GPU monitors, and a terminal text editor.

BashPythonDHCPLinux

Habit Tracker

Flutter habit tracker — custom habits, reminders, streak visualisation and progress insights over time.

FlutterDartMobile

Notes App

Flutter note taking with text, image and audio capture, organised by tags, categories and notebooks, with search and filtering across notes.

FlutterDartMobile

Shoe Store

Flutter storefront — search by brand, style, colour, size and price, product detail with reviews and ratings, cart and checkout, plus order history with status tracking.

FlutterDartE-commerce

04 Stack

Tools I reach for, grouped by what they solve.

LLM & Agentic AI

LangChainLangGraphRAG Semantic chunkingHybrid vector + lexical Multi-agent systemsTool calling Prompt engineeringOllamaLoRA / QLoRA

Backend & MCP

FastAPIPython (asyncio)FastMCP MCP server designSSE transport REST API designNode.jsExpress.js

Data & Storage

PostgreSQLpgvectorPostgres full-text (tsv) MinIO / S3 (aioboto3)chromadbMongoDBMySQL FHIR R4

Auth & Security

Firebase AuthJWTOAuth 2.0 Fernet encryptionRBACbcrypt

Languages

PythonJavaScriptCC++JavaDart

Frameworks & Tools

pytestruff / blackDockerAlembic SQLAlchemypdfplumberreportlab Tkintermatplotlib FlutterReactGitLinux (Arch)

05 Credentials

Education and certifications.

Education

2022 — 2026

B.Sc. in Computer Science & Engineering

Daffodil International University

CGPA 3.72 · all coursework completed

2019 — 2021

HSC, Science

Gurudayal Govt. College

GPA 5.00 / 5.00

Certifications & competitions

  • AI
    AI+ Prompt Engineer Level 1™ AI CERTs® · Jun 2025
  • AX
    Agent X Certified NetCom Learning® / AI CERTs · Sep 2025
  • 5D
    5-Day AI Agents Intensive Google & Kaggle · Nov 2025
  • ML
    Workshop: AI-Powered Learning Processing DIU NLP & ML Research Lab · Nov 2025
  • DT
    Datathon — CUET CSE Fest 2025 Team LazyPenguins

06 Contact

Building something agentic? Let's talk.