8000 GitHub - Mahdi-Hasan/Mahdi-Hasan · GitHub
[go: up one dir, main page]

Skip to content

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

1 watching

Forks

Latest commit

 

History

57 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Mehedi Hasan · Full-Stack Software & Product Engineer · Senior Software Engineer @ CloudApper AI · 5+ years · 74 open-source PRs, 31 merged, 23 projects

Merged PRs to microsoft/agent-framework, HKUDS/DeepTutor and ivmartel/dwv

LinkedIn Email Hugging Face Resume

Previously Software Engineer at Vivasoft

Open source

Hero counts refresh daily from GitHub. PR badges update live.

⭐ Featured PRs

microsoft agent-framework stars
Microsoft's agent SDK · Python

3 merged fixes to model clients. A boto3 session set to eu-west-1 was sent to us-east-1 with no error. Mistral and Gemini now respect an injected SDK client.

#9967 #9968 #9969
vercel next.js stars
Vercel's React framework · TypeScript

App Router-only builds returned 500 for client errors Next.js raised itself, such as a bad param or a bad Range. They now keep the 4xx status.

#101188
openclaw openclaw stars
AI assistant · 390K+ stars · TypeScript

4 fixes across IRC, LINE, formatting and the dashboard. A fallback IRC nick grew past the 30-character limit. Sizes like "1024 KiB" now roll over to "1.0 MiB".

#196619 #196631 #196636 #196639
JetBrains thinkrail stars
JetBrains' agentic IDE · TypeScript · WSL

Editor lookup scanned a slow WSL PATH on the event loop and froze the host for seconds. "Reveal in file manager" opened a hidden Explorer window on Windows.

#1158 #1159
Nano-Collective nanocoder stars
Open coding agent · TypeScript

2 merged fixes: MCP tool results kept only the first content block, and string_replace missed matches in CRLF files. Also an XSS fix in the VS Code webview.

#1960 #1961 #1927
HKUDS DeepTutor stars
AI tutor · Python · LLMs

3 merged LLM fixes. Reasoning-model replies broke lesson JSON. Strict gateways rejected replayed items. Long Claude Code prompts failed on Windows.

#2375 #2454 #2456

📂 All contributions

🤖 AI agents & LLM frameworks · agent-framework, openclaw, DeepTutor, crewAI, pydantic-ai, LangChain, open-notebook
Project What I fixed PR
microsoft/agent-framework
stars
Python boto3 Gemini SDK pytest
Bedrock: a boto3 session set to eu-west-1 was sent to us-east-1 when BEDROCK_REGION said so. No error was raised. The session's region now wins. #9967
Mistral: service_url now reports the endpoint of the injected SDK client, not the default URL. #9968
Gemini: with an injected genai.Client, a bad env value such as GOOGLE_GENAI_USE_ENTERPRISE still crashed construction. Those settings are now skipped. #9969
Foundry hosting: cumulative usage snapshots were added together, so 100 then 250 input tokens showed as 350. Each agent's snapshot now replaces only that agent's usage. #10205
openclaw/openclaw
stars
TypeScript Node.js Vitest
IRC: after repeated nick collisions the fallback nick grew past the 30-character limit. It is now cut to fit. #196619
Formatting: sizes just below a unit boundary showed as "1024 KiB". They now roll over to "1.0 MiB". #196631
LINE: email links in table cells showed twice, as addr (mailto:addr). They now show once. #196636
Dashboard: author search bolded the wrong letters for Korean and accented names. Highlights now match the typed text. #196639
HKUDS/DeepTutor
stars
Python LLMs
Lesson prep broke on reasoning models when <think> blocks or prose wrapped the JSON. It now finds the JSON and retries once at the right effort level. #2375
Responses API: strict gateways rejected status on replayed items such as function_call. It is now stripped from every item. #2454
Workspace: a failed probe-file delete escaped the request. Cleanup is now best effort. #2455
Claude Code: system prompts over 32K characters failed on Windows with a misleading "not found" error. Long prompts now go through a file. #2456
crewAIInc/crewAI
stars
Python pytest
Files: expired-upload cleanup never called the provider's delete, because the lookup had already dropped expired entries. Expired files are now deleted. #9109
Telemetry: strings, numbers and None at the depth limit became repr() strings. Primitives are now kept unchanged. #9110
A2A auth: a comma inside a quoted WWW-Authenticate value split the challenge. Parsing now respects quotes and escapes. #9111
Planning: replanned steps restarted at 1 and clashed with completed steps. Clashing steps are now renumbered. #9136
CLI: after an upgrade, a cached yanked flag from the old version showed a false warning. The cache is now tied to the installed version. #9137
Executor: every batch after a replan asked for another replan over failures already handled. Handled failures are now skipped. #9138
pydantic/pydantic-ai
stars
Python asyncio SQLite pytest
OpenAI batch: grouping read default_headers, which ran uname -p on the event loop. It now hashes only the auth, org and project headers. #12590
Tests: the Mistral max_retries test made one attempt on slow runners. It now uses a fake clock, so it passes at any speed. #12591
Planning: SqlitePlanStore leaked its connection when schema setup failed on a locked database. It now closes and re-raises. #12592
OpenMontageInc/OpenMontage
stars
Python ComfyUI FFmpeg
The ComfyUI video tool reported request defaults as output metadata. It now reports the real values probed from the file. #885
tryAGI/LangChain
stars
C# .NET RAG
AddDocumentsFrom* helpers now take several data sources in one call for RAG ingestion. #639
lfnovo/open-notebook
stars
Python SurrealDB
Sources: a delete that hit a SurrealDB conflict returned 200, but the source and its chunks stayed. Errors now raise, and conflicts are retried. #1559
Sources: a NUL byte in a text source left a broken command row that stopped the worker. NUL bytes are now rejected before the row is written. #1564
🩺 Healthcare & medical imaging · MONAI, OHIF, PyHealth, dwv, OpenEMR
Project What I fixed PR
Project-MONAI/MONAI
stars
Python PyTorch pandas
Data loading: CSVIterableDataset with shuffle=True returned every item twice per epoch on stream sources. The unshuffled pass now runs only when shuffle is off. Added a regression test. #9249
OHIF/Viewers
stars
TypeScript React Jest DICOM
DICOM viewer: the segment stats card kept stale L/W values and eye state until re-hover. The React Compiler cached the service reads. The card now holds them in state and refreshes on SEGMENTATION_MODIFIED. #6408
Docs: the FocalLoss alpha docstrings did not describe per-class weights. They now match the merged behaviour. #9251
Scrollbar: an inactive viewport's scrollbar kept a white focus ring. It now takes focus only when keyboard navigation is on. #6409
sunlabuiuc/PyHealth
stars
Python ML EHR
StageNet: nested visits longer than the fitted width were not truncated correctly. The processor now logs one warning per processor. #1380
ivmartel/dwv
stars
JavaScript DICOM
Charset: Korean DICOM strings coded as ISO 2022 IR 149 failed to decode. They now decode as EUC-KR. #2370
Performance: canvases read back with getImageData now use willReadFrequently, mainly for video frames. #2373
Tests: added a regression test for sequences that mix defined and undefined length items. #2374
Loading: URL loads take new timeout and timeoutRetries options. #2375
Units: lengths from raw Imager Pixel Spacing now read mm (detector), not plain mm. #2376
openemr/openemr
stars
PHP EHR
Errors: HTTP error responses carried no useful body. They now send the status text, and exception detail only when error display is on. #15026
🛠️ Developer tools · orca, thinkrail, nanocoder, next.js, claude-buddy, Crow
Project What I fixed PR
stablyai/orca
stars
TypeScript Electron React Zustand
Sidebar: a new bindable command collapses or expands every project at once. Before, each chevron had to be clicked. #37718
Jira: several connected sites all showed the same profile name. Each site now shows its host, like Ada (acme.atlassian.net). #37719
Mobile: "Jump back in" rows showed only the branch, so two main rows looked the same. The repo name now shows on every row. #37720
Browser: an opt-in setting loads terminal links in the open browser tab instead of a new split pane. #37721
JetBrains/thinkrail
stars
TypeScript Node.js WSL
Windows: "Reveal in file manager" opened Explorer in a hidden window, with no error. The folder window now shows. #1158
Performance: editor lookup scanned a slow WSL PATH on the event loop and froze the host for seconds. It now pre-filters folders with async readdir. #1159
Nano-Collective/nanocoder
stars
TypeScript VS Code API DOMPurify
Security: the VS Code chat webview now sanitizes markdown and blocks remote images. This closes an XSS and tracking-pixel path. #1927
MCP: only the first content block of a tool result was read. Every block is now formatted, in order. #1960
Edit tool: string_replace missed matches in CRLF files. It now matches and keeps CRLF line endings. #1961
Write tool: write_file turned a CRLF file into LF when it overwrote it. Line endings of CRLF files are now kept. #1992
Edit tool: diff_edit missed matches in CRLF files. It now matches and keeps CRLF line endings. #2007
vercel/next.js
stars
TypeScript Node.js
App Router-only builds returned 500 for client errors Next.js raised itself, such as a bad param or a bad Range. They now keep the 4xx status. #101188
handsome-rich/claude-buddy
JavaScript Electron PowerShell
10 PRs for a Claude Code desktop companion. A floating pet with status badges, tool-permission allow/deny, Windows autostart and a packaged-exe fix. #3–#13
Pienteger/Crow
C# .NET PowerShell
4 PRs for a build-cleanup CLI. --ig now protects removal targets, paths with spaces work, test teardown no longer fails, and a one-line installer. ✅ #2–#5 merged
🕌 Community & Islamic tech · quran.com, QUL, hadith-api, DeenTime, Ansari, Mihrab
Project What I fixed PR
quran/quran.com-frontend-next
stars
TypeScript Next.js React
Security: .env.example turned off TLS checks by default. New setups now keep TLS verification on. #3362
Dates: across a DST change, relative dates showed 4.042 days ago. The day count is now rounded, and the caller's Date objects are no longer changed. #3368
Locale: a number-format call with custom options switched the cached locale, so English pages could show Arabic digits. The cache is now keyed by locale. #3369
Reflections: a plain URL near an <a> tag, or a repeat of a linked URL, stayed plain text. Each match is now checked at its own position. #3370
DeenTime
JavaScript adhan-js
The prayer-time engine gave NaN or times hours off. I restored the solar math and checked it against adhan-js. #69
Fixed three tests that failed on main. #70
a11y: calendar month buttons now hide decorative icons, name the grid they control and announce the new month. #72
a11y: the 99 Names dialog now closes with Escape in browsers without showModal(). #73
Search: 99 Names search now ignores diacritics and Arabic harakat. #74
a11y: the Tasbih counter announces when the target is reached. #75
Reset: the Tasbih display now refreshes after settings are reset. #76
TarteelAI/quranic-universal-library
stars
Ruby JavaScript Docs
Docs: the Quran Metadata tutorial assumed arrays and fields no export has, and threw surahs.map is not a function. Snippets now match the real export. #1021
Docs: the Surah Information snippet used old field names and rendered an empty card. It now reads the fields the exporter writes. #1022
Docs: the translation snippet threw on chunked exports. It now handles every export shape, including footnotes and SQLite rows. #1023
fawazahmed0/hadith-api
stars
JavaScript Data
Text cleanup: cleanTrans cut closing quotes, full stops and reference numbers from every hadith. It now strips only stray symbols. #144
Data: eng-bukhari and eng-malik rebuilt with the fixed cleanup, restoring trimmed endings without changing any hadith text or number. #146
Data: fixed a broken >br> token in eng-muslim hadith 4601. #145
Ansari
TypeScript Node.js PostgreSQL
The API's Postgres pool had no error listener. A dropped idle connection became an uncaught exception. It is now logged and the dead client is discarded. #349
Mihrab
TypeScript React Native Kotlin
Arabic home-screen widgets were not RTL. I traced it to widget rows following the phone's locale, not the app's language. The maintainer shipped a fix based on it. #118

About me

I'm a full-stack software & product engineer. I start from the business problem, not the tool. Then I ship the fix end to end: data, back end, AI and UI.

  • 🎯 How I work: find where the workflow hurts, pick the simplest tech, measure the result.
  • 🏗️ Full stack: 5+ years of C# / .NET and Python, with React, Angular and Vue front ends.
  • 🌍 Open source: 74 PRs, including Microsoft, JetBrains, Vercel and healthcare AI (MONAI, OHIF, OpenEMR, dwv, PyHealth).
  • 🧭 Product: certified AI-native Product Manager (Pendo). I work AI-native with Claude Code.
  • 🎓 Teaching: mentor for the World Bank-backed Edge Program. Also lead small teams and mentor juniors.

🤖 AI I've shipped

🧩 Agentic copilots
Build apps, forms & workflows from chat · MCP server + client
🧪 Code agents
LangGraph · runs code in gVisor / Firecracker sandboxes
📚 RAG & memory
Embeddings · semantic search · knowledge-graph memory
🔀 Multi-model routing
LiteLLM · OpenAI + Claude · reasoner & verifier models
🛡️ Safety & evals
Guardrails · approval gates · replay tests · LLM-as-a-Judge
💸 Cost & context
Prompt caching · context compaction · per-call cost tracking
📄 Document & media AI
PDF & image vision · field extraction · transcription
🏛️ Legacy modernization
Agents read COBOL, PL/I, RPG, VB6, Delphi → rebuild specs

🏢 Domains

HR & Payroll Healthcare, HIPAA Finance Ops No-code Platforms Legacy Modernization Identity & Security

Impact in numbers

60K+ SSO users · 800+ companies on PIHR · ~100K invoices at 90% accuracy · 70%+ less manual work · 20–30% faster APIs · zero PHI leaks · ~2× Bangla summarization accuracy · Top 5% LinkedIn C# & OOP

Experience

CloudApper AI (a division of M2SYS Technology) — Senior Software Engineer · Jun 2026 – Present
  • Built an AI app-builder copilot. It plans and builds apps, forms, workflows and dashboards from a chat, then checks its own work. Serves its tools over MCP and calls external MCP servers with per-user auth. (TypeScript, Fastify, Angular, Postgres, Redis)
  • Built a sandboxed code agent on LangGraph. It writes and runs code in gVisor / Firecracker sandboxes, with human approval, web search and resumable runs. (Python, FastAPI, Angular)
  • Set up multi-model routing through a LiteLLM gateway across OpenAI and Claude. Reasoner and verifier models review risky changes. Prompt caching, context compaction and per-call cost tracking keep spend in check.
  • Building a dynamic agent engine in Python on the Claude Agent SDK, LangGraph and LiteLLM. It turns a plain instruction into an agent with its own skills, tools and code sandbox. A router agent picks the simplest path.
  • Added per-user shared memory across the Marketing, Lead Gen and Sales agents. Customer context carries from one agent to the next.
  • Shipped a run-deliverables API with SSE streaming and REST retrieval. Users see live progress, and dropped streams recover cleanly.
  • Moved agent permissions into the tool layer: pre-scoped tool wrappers and human approval on destructive actions. Also blast-radius limits, self-evaluation, policy guardrails and replay tests.
  • Built AI connectors for the no-code platform: RAG knowledge bases, semantic search, field extraction, media transcription and a code-agent plugin. (.NET)
  • Architected Legacy → CloudApper Transform. A Director agent, 4 domain Supervisors and specialist agents read COBOL, PL/I, RPG, VB6 and Delphi code. They produce validated SRS documents. (.NET 10, Python, Kafka, LiteLLM, GCP)
  • Red-teamed a multi-tenant MCP platform for prompt injection, tool-result poisoning and cross-tenant access. Findings led to sandbox isolation, read-only DB fallback, query sandboxing and workload identity.
  • Designed A2A agent swarms and serverless agents such as OpenClaw for HIPAA, SOC 2 and GDPR customers. Agents spawn with least-privilege tools, retry on failure, then hand off to a human.
Vivasoft Limited — Software Engineer, Level 2 · Aug 2021 – May 2026 (4 yrs 10 mos)
  • Built AI agent microservices with LLMs, RAG, auto SQL generation and vector DBs. They plug into C# .NET Core platforms via gRPC.
  • Engineered .NET Core microservices (SOA, Clean Architecture, CQRS) for healthcare and HR platforms.
  • Built Python FastAPI services and Azure Functions ETL pipelines. They turn messy documents and third-party data into clean records, with retries and clear errors.
  • Automated a workflow across several services, cutting manual effort by over 70%.
  • Built a central SSO service with Redis for 60K+ users.
  • Built eval and test setups: golden datasets for LLM evals, k6 load tests, Grafana dashboards.
  • Cut latency ~20% with Redis caching and SQL tuning. RabbitMQ background jobs gave 30% faster responses.
  • Ran CI/CD, Docker and Kubernetes (ACA/AKS) with least-privilege IAM.
  • Led small project teams and mentored juniors on agent orchestration and back-end design.

Featured projects

☁️ At CloudApper AI

🏛️ Legacy → CloudApper Transform

Multi-agent code modernization

  • A Director agent leads 4 domain Supervisors and specialist agents.
  • Reads legacy COBOL, PL/I, RPG, VB6 and Delphi codebases.
  • Produces validated SRS and requirement models for the rebuild.
  • I surveyed DeepWiki, Sourcegraph Cody, AWS Transform, IBM ADDI and EPAM ART to shape the design.

.NET 10 Python Kafka LiteLLM GCP

🛡️ Secure multi-tenant MCP agent platform

Agent safety · red-teaming

  • Red-teamed for prompt injection, tool-result poisoning and cross-tenant access.
  • Findings led to sandbox isolation and approval gates on destructive tool calls.
  • Read-only DB fallback, query sandboxing and workload identity federation.
  • Agent code runs in gVisor containers with allowlisted tools.

MCP gVisor Microsoft Agent Framework GCP

🤖 Dynamic agent engine & A2A swarm

Agentic workflows · HIPAA, SOC 2, GDPR

  • Turns a plain instruction into an agent with its own skills, tools and code sandbox.
  • A router agent picks the simplest path for each task.
  • Per-user shared memory links the Marketing, Lead Gen and Sales agents.
  • Agents work together over A2A. Failed tasks retry, then go to a human.

Claude Agent SDK LangGraph LiteLLM A2A MCP Python

📑 Enterprise document pipeline

Human-in-the-loop extraction

  • Gemini-based extraction of fields from enterprise documents.
  • Human review gate before any result is saved.

Gemini Python GCP

🧪 Earlier work & side projects

🩺 Sybil — Multi-agent healthcare AI

HIPAA-regulated · production

  • 6 agents with LangGraph orchestration and MCP tools, running on GKE.
  • LiteLLM gateway for routing, fallback and caching. Hybrid RAG on OpenSearch.
  • PHI scanners, calibrated LLM-as-a-Judge and canary rollouts gated by stats tests.
  • Every write goes to an RBAC Postgres audit store. Zero PHI leaks.
  • Document parsing at 90% accuracy on ~100K invoices from 3,000+ clients.

Python LangGraph MCP .NET Azure Doc Intelligence

🧠 Self-Orbit — Personal intelligence platform

Open source

  • .NET 10 services (Clean Architecture, CQRS, DDD) behind a YARP gateway.
  • Python FastAPI AI engine for embeddings and LLM inference, linked over gRPC.
  • Next.js 15 front end in strict TypeScript, with server-first rendering.
  • Turns text and audio into tasks, expenses and summaries.
  • Unit, integration and contract tests (xUnit, Moq, pytest) in an Nx monorepo.

C# Python gRPC Next.js 15 Kubernetes

🏢 PIHR — SaaS HR & payroll

Enterprise · 800+ companies

  • Payroll, attendance and leave microservices for 50,000+ users.
  • RabbitMQ and Redis on AWS. EF Core and Dapper tuning cut query time ~25%.
  • RAG document search, validated with golden datasets.

C# ASP.NET Core RabbitMQ Redis AWS

Open NLP research

  • Fine-tuned mT5-large on a large Bangla corpus.
  • ~2× accuracy over earlier Bangla summarization models.
  • Model is public on Hugging Face.

Python PyTorch Transformers mT5

  • Ranks all 64 districts by temperature and PM2.5 air quality.
  • <500 ms p99 from a pre-computed Redis cache with background sync.
  • Clean Architecture and CQRS, with unit, integration and E2E tests.

.NET 10 Redis Hangfire xUnit

  • An MCP server to query Azure Application Insights in plain English.
  • Ask "show me errors from the last 24 hours" inside Claude.

Node.js MCP Azure

📘 SaaS-based AI guide

  • RAG product that turns uploaded documents into searchable product guides.
  • Parses uploads, embeds them with Jina and stores them in a vector DB and MongoDB.
  • Source files stay on Amazon S3.

Python FastAPI LangChain MongoDB S3

Tech stack

Languages
Languages

Back end & front end
Back end and front end

Data, AI & ML
Data and AI

Cloud & DevOps
Cloud and DevOps

Tools
Tools

Full tool list
Area Tools
🧑‍💻 AI-native workflow Claude Code, custom Claude skills, MCP servers, AGENTS.md-driven repos, AI-assisted specs and code review
🤖 Agent frameworks Claude Agent SDK, Microsoft Agent Framework, OpenClaw (serverless agents), LangGraph (checkpointers, interrupts), LangChain (Python & C#), OpenAI SDKs (Python, .NET), MCP TypeScript SDK
🔌 Agent protocols MCP (servers, clients, per-user auth, tool tiers, deferred tool loading), A2A
🧠 LLMs OpenAI, Azure OpenAI, Claude, Gemini, Hugging Face Transformers, mT5, PyTorch
⚙️ LLM ops LiteLLM gateway (routing, fallback, caching), multi-model tiers (reasoner, verifier), prompt caching, context compaction, token budgets, cost tracking, SSE streaming
🏃 Agent runtime gVisor and Firecracker sandboxes, E2B, tool calling (parallel), human-in-the-loop approvals, resumable runs, thread memory, knowledge graphs, web search (Tavily)
🔎 RAG & search Embeddings (Jina, OpenAI), hybrid search, semantic search, OpenSearch, Solr, Milvus, vector DBs, Neo4j graph DB, Postgres full-text search, auto SQL generation
📄 Document & media AI Azure Document Intelligence, OCR, Gemini extraction, PDF and image vision, Office file parsing, audio and video transcription, structured output
🛡️ AI safety & evals Approval gates, least-privilege tools, policy guardrails, red-teaming, PHI scanners, LLM-as-a-Judge, self-evaluation, replay tests, golden datasets
📡 Observability OpenTelemetry, Langfuse, Azure App Insights, Grafana
🟣 .NET ASP.NET Core, Web API, EF Core, Dapper, gRPC, YARP, Hangfire, plugin-based connectors
🐍 Python FastAPI, Pydantic v2, LangGraph, SQLite, pytest, pytest-asyncio, Jupyter
🟩 Node.js TypeScript, Fastify, Zod, pnpm, Turborepo, Vitest
🎨 Front end React, Next.js, Angular, Vue, TypeScript, Tailwind CSS, Framer Motion, Radix UI, Nx monorepos
📨 Messaging RabbitMQ, Kafka, event-driven systems
🔗 Integrations Webhooks, ETL batch and stream pipelines, partner data ingestion, Gmail API, no-code plugin connectors
🗄️ Databases PostgreSQL, SQL Server, SQLite, MongoDB, Redis, Upstash
☁️ Cloud Azure (ACA, AKS, Functions, App Insights), AWS (S3), GCP (GKE), Vercel
🚢 DevOps Docker, Kubernetes, Terraform, GitHub Actions, GitLab CI, Jenkins
🏛️ Architecture Microservices, SOA, Clean Architecture, CQRS, DDD, event-driven design
🧪 Testing xUnit, NUnit, Moq, TDD, pytest, Vitest, k6 load testing, contract tests, LLM replay tests
📜 Governance OWASP Top 10 for Agentic Apps, NIST AI RMF, EU AI Act, HIPAA, SOC 2, GDPR

Recognition & learning

Let's talk

Open to AI product engineering, agentic AI and full-stack roles.

LinkedIn Email

Profile views

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages

0