Skip to content
Thaher., back to the top
Available for software engineering roles

Mohammed Thaher S

I build and ship production-ready full-stack systems, Hybrid ReAct agents, RAG pipelines and multimodal AI apps

CS graduate specializing in custom Hybrid ReAct architectures, RAG pipelines and end-to-end multimodal applications, with LLM agents embedded straight into products served via FastAPI and Django. Every project is a working system with a live demo.

View my workGet in touchResume(opens in a new tab)
  • Chennai, India
  • Open to relocate anywhere in India
  • Available for remote roles
Live AI systems
4
Each one deployed with a public demo
Django views integrated
30+
TravelKeet launch, IIT Madras Research Park
Crop-disease accuracy
95%
Winner, AgriTech Hackathon
ML evaluation score
94.14%
Finalist, Zelestra x AWS ML Hackathon
  • Python
  • FastAPI
  • Django
  • LangGraph
  • LangChain
  • Groq
  • Gemini
  • FAISS
  • YOLOv8
  • OpenCV
  • React
  • Flask
  • PostgreSQL
  • MongoDB
  • Docker
  • Pytest
  • Streamlit
  • Gradio

01About

Every project ships as a working system.

I am a CS graduate who builds and ships production-ready full-stack and AI-integrated systems. I specialize in designing custom Hybrid ReAct architectures, RAG pipelines, and end-to-end multimodal applications.

I handle frontend-to-backend integration, directly embedding LLM agents and workflows into products served via FastAPI and Django. Every project I build is a working system with a live demo.

At a glance

Based in
Chennai, India
Relocation
Open to anywhere in India
Remote work
Available for remote roles
Education
B.E. Computer Science, 2026
Honors
First Class with Distinction

  • Agentic systems

    Hybrid ReAct loops and LangGraph state machines, with deterministic guardrails where the business logic demands them.

  • Retrieval pipelines

    PDF ingestion, FAISS vector search and cited answers, with relevance routing for multi-step reasoning.

  • Multimodal AI

    Computer vision and language models in one flow, with YOLOv8 detections handed to Gemini for scene reasoning.

  • Production backends

    Async FastAPI and Django services with Pydantic validation, Pytest suites and clean frontend integration.

Languages I speak

Six languages, three of them native.

  • TamilNative
  • EnglishFull professional
  • UrduNative
  • HindiNative
  • GermanBasic
  • ArabicReading

02Selected work

Engineering & AI systems

End-to-end architectures and self-reflecting agents. Every one of them is deployed and live.

Live demos are hosted on free-tier infrastructure. Please allow 30 to 60 seconds for cold starts outside standard IST business hours.

trendly/architectureSystem design
RequestFastAPIReAct looppython-nativeThoughtActionObserveiterateGroq LLMLlama 3runs in parallelRules enginedeterministic8 business-logic guardrailse.g. return window, PIIResponsePydantic
01

Trendly

A Hybrid ReAct agent with a deterministic guardrail engine.

I designed a custom Hybrid ReAct architecture from scratch featuring a Python-native ReAct loop governing a Groq LLM. It includes a parallel deterministic rules engine enforcing 8 business-logic guardrails, such as return window expiry and PII protection. I served the application via async FastAPI with Pydantic validation and implemented comprehensive Pytest suites.

  • Python-native ReAct loop built from scratch
  • 8 deterministic business-logic guardrails
  • Async FastAPI with Pydantic validation and Pytest suites
  • FastAPI
  • Groq (Llama 3)
  • Pydantic
  • Pytest
Open live demoof Trendly (opens in a new tab)trendly-agent-1.onrender.com/ui
documind/architectureSystem design
STARTplan7 doc typesexecuteper sectionreflectself-critiquereviseapply fixesexport.docxENDflags timeline inconsistenciesand hallucinated figuresLangGraph state machine · 5 nodes
02

DocuMind

A self-reflecting document generator built on LangGraph.

I engineered a LangGraph state machine with 5 nodes (plan, execute, reflect, revise, export) that classifies 7 document types and generates sections individually. The pipeline runs a self-critique pass that flags timeline inconsistencies and hallucinated figures before exporting a polished .docx file. The React frontend is served from the same async FastAPI process.

  • 5-node LangGraph state machine
  • Classifies 7 document types, writes section by section
  • Self-critique flags timeline and figure errors before export
  • LangGraph
  • FastAPI
  • React
Open live demoof DocuMind (opens in a new tab)docu-mind.up.railway.app
multimodal-visual-qa/architectureSystem design
CV inferenceLLM reasoningpersonbicycleimageYOLOv8m80 COCO classesSpatial contextlabels + boxesQuestionopen-endedGemini 1.5 Flashvia LangChainAnswerscene reasoningasync FastAPI backendnon-blocking concurrent requests
03

Multimodal Visual Q&A

Open-ended scene reasoning on top of object detection.

I built a two-stage pipeline where YOLOv8m extracts spatial metadata across 80 COCO classes, and this structured context is injected into Gemini 1.5 Flash via LangChain for open-ended scene reasoning. CV inference and LLM reasoning are decoupled services behind an async FastAPI backend, deployed with async I/O for non-blocking concurrent requests.

  • YOLOv8m spatial metadata across 80 COCO classes
  • Gemini 1.5 Flash reasoning via LangChain
  • Decoupled CV and LLM services with non-blocking async I/O
  • YOLOv8
  • Gemini Flash
  • LangChain
  • FastAPI
Open live demoof Multimodal Visual Q&A (opens in a new tab)7haher-multimodal-vision-qa.hf.space
doc-rag-chatbot/architectureSystem design
PDFChunk & embedingestionFAISSvector storeQuestionretrievetoolrelevant?yesnogenerateLlama 3 · GroqAnswercitedweb searchtoolLangGraph agent · 3 tool nodes · relevance routing
04

Doc RAG Chatbot

Cited answers from your PDFs, with relevance-routed web search.

I developed a full RAG pipeline with PDF ingestion, a FAISS vector store, and cited answer generation via Llama3 on Groq. I extended this into a LangGraph autonomous agent with 3 tool nodes (retrieve, web search, generate) and a conditional relevance-routing edge for multi-step reasoning.

  • PDF ingestion into a FAISS vector store
  • Cited answers generated by Llama 3 on Groq
  • LangGraph agent with 3 tool nodes and relevance routing
  • LangChain
  • LangGraph
  • FAISS
  • Groq (Llama 3)
  • Streamlit
Open live demoof Doc RAG Chatbot (opens in a new tab)doc-rag-chat-bot.streamlit.app

03Experience

Experience & education

Shipping production code inside a team, backed by a computer science degree with distinction.

  1. Work

    Backend Developer Intern

    Campervan Factory · IIT Madras Research Park

    to 4 months
    • I integrated frontend UI components with a Django backend for the TravelKeet platform launch, ensuring smooth data flow and session management across 30+ views.
    • I collaborated within a cross-functional team using Git for version control and coordinated delivery, while managing and mentoring 2 fellow interns.
    views integrated
    30+
    interns mentored
    2
    • Django
    • Session management
    • Git
    • Mentoring

  2. Education

    B.E. Computer Science and Engineering

    Sathyabama Institute of Science and Technology

    to 4 years
    • Graduated First Class with Distinction.
    with Distinction
    First Class

04Skills

Skills & tech arsenal

The languages, frameworks and tools behind every system on this page.

  • Languages

    Python first, with the web stack and systems fundamentals.

    06
    • Python (primary language)
    • JavaScript
    • HTML
    • CSS
    • Java
    • C

  • AI & ML

    Agents, retrieval and computer vision.

    07
    • LangChain
    • LangGraph
    • FAISS
    • YOLOv8
    • OpenCV
    • Groq
    • Gemini

  • Backend & APIs

    Async services and the data stores behind them.

    05
    • FastAPI
    • Django
    • Flask
    • PostgreSQL
    • MongoDB

  • Infra & Tools

    Packaging, testing and demo tooling.

    05
    • Docker
    • Git
    • Pytest
    • Streamlit
    • Gradio

05Hackathons

Competitive engineering

Models trained, tuned and shipped against a competition clock.

  • Winner
    95%classification accuracy

    AgriTech Hackathon

    Pavaman Technologies

    I fine-tuned EfficientNetB4 for 38-class crop disease detection, achieving 95% classification accuracy. I shipped a live Streamlit demo under the competition deadline.

    • EfficientNetB4
    • 38 classes
    • Streamlit
    View live demofor AgriTech Hackathon (opens in a new tab)
  • Finalist
    94.14%evaluation score

    Zelestra x AWS ML Hackathon

    HackerEarth

    I built a stacking ensemble utilizing CatBoost, LightGBM, TabNet, and XGBoost with Optuna HPO over 50 trials. I achieved a 94.14% evaluation score with SHAP interpretation.

    • CatBoost
    • LightGBM
    • TabNet
    • XGBoost
    • Optuna
    • SHAP
    View notebookfor Zelestra x AWS ML Hackathon (opens in a new tab)

06Contact

Let's build something that ships.

I am available for software engineering roles. Based in Chennai, open to relocating anywhere in India, and available for remote work.

Email

thahercareer@gmail.com
Open mail appResume(opens in a new tab)

Chennai, India (IST)

Send a message

Roles, projects or a question about one of the systems above. It goes straight to my inbox.

What's this about?

Your details are only used to reply to you.

Made with Modulify