Keshav Kahar, AI Engineer & Software Development Engineer. I build AI features into software products.

At CaseFox I work on AI for legal documents: retrieval pipelines, LLM workflows, and the FastAPI services that integrate them with the application. Underneath that is 3+ years of software development in Python, from backend APIs to computer-vision pipelines.

See the projects Get in touch Resume (opens in a new tab)
An embedding field A scatter of points representing topics from Keshav's work, such as RAG, vector search, prompt design, NL to SQL, FastAPI, legal drafting and meeting notes. A highlighted query point lights up and connects to the topics nearest to it. legal research semantic search legal drafting RAG embeddings doc intelligence vector search summarization source context FAISS · Chroma LLMs prompt design MySQL meeting notes NL to SQL LangChain action items FastAPI Python insights REST APIs OpenCV query
Fig. 1 Retrieval, drawn. A query lights up its nearest neighbors in a field of topics from my work. Move your pointer to ask something else.

Five questions I answer before writing a prompt.

Calling a model is rarely the hard part of an AI feature. These questions decide the design, and each answer points to a role or project where I worked that way.

  1. “Where does the answer actually live?”

    Match the mechanism to the data.

    Text in documents calls for retrieval. Rows in tables call for a generated SQL query that the database runs. A conversation calls for extraction into fields. I have built all three, and choosing between them comes first.

    Where I did this 1Legal Drafting & Summarization 2Database Search 4Meeting Intelligence

  2. “What may the model write from?”

    Retrieved source text, not memory.

    For drafting, summarizing and research, the pipeline first fetches the relevant source context and passes it to the model with the request. Embeddings, a vector database such as FAISS or Chroma, and semantic search do the fetching.

    Where I did this 1Legal Drafting & Summarization 3EDGAR Legal Research CaseFox role

  3. “Who, or what, reads the output?”

    Shape the output for its consumer.

    A person can read prose; software cannot use it. When a follow-up workflow consumes the result, the pipeline returns separate structured outputs, such as a summary, insights and action items, instead of one block of text.

    Where I did this 4Meeting Intelligence

  4. “How does the rest of the product call it?”

    Put it behind an API.

    An LLM workflow becomes a feature once other code can call it. I design FastAPI services and REST endpoints so the application integrates with the workflow like any other backend service.

    Where I did this CaseFox role

  5. “Where will it be slow?”

    Treat it as backend software.

    A pipeline with a model in it still has ordinary backend problems. Finding them is work I did at Zomit: diagnosing performance bottlenecks and optimizing backend processing and data workflows.

    Where I did this Zomit role

Four projects. Four different engineering problems.

Each one is described the same way: the problem, what I built, how a request moves through it, and what it delivers.

The first was built in my role at CaseFox, so it is described at the architecture level.

4Structured extraction · Workflow

Sales Meeting Intelligence System

An LLM pipeline that turns a sales conversation into a structured summary, insights and action items.

  • LLMs
  • NLP
  • Structured output
  1. Conversation data
  2. LLM pipeline
  3. Summary, insights, action items
  4. Follow-up workflow
Fig. 5One conversation, three structured outputs.

Problem

What is said in a sales conversation matters after it ends, but most of it survives only as scattered notes. Follow-up depends on what someone remembered to write down.

What I built

I developed the LLM pipeline that processes conversations into structured summaries, insights and action items.

How it works

The pipeline reads the conversation data and returns three separate structured outputs instead of one block of prose, so a follow-up workflow can use each part directly.

What it delivers

For each conversation: a summary, a list of insights and a list of action items, in the same structure every time.

Experience

Backend engineering first. AI engineering built on it.

I started as a Python developer writing automation scripts, moved on to backend APIs and computer-vision pipelines, and now build LLM features at CaseFox. The AI work rests on the software engineering underneath it.

  1. Now · AI engineering

    Software Development Engineer

    CaseFox to present

    Rising Star Award, 2025

    • Engineered RAG pipelines that draft and summarize legal documents, combining LLM generation with retrieved source context.
    • Developed an AI assistant for context-aware information retrieval using LangChain, embeddings, vector databases and semantic search.
    • Designed FastAPI services and REST endpoints that integrate LLM workflows with backend applications.
    • Built an NLP meeting-intelligence pipeline that turns conversation data into summaries, insights and follow-up actions.
    • RAG
    • LangChain
    • Embeddings
    • Vector databases
    • FastAPI
    • REST APIs
    • NLP

  2. Before · Backend engineering

    Software Developer, AI & Automation

    Zomit to

    • Developed Python backend APIs that support AI application and process-automation workflows.
    • Integrated OpenCV computer-vision pipelines for real-time data processing inside application services.
    • Diagnosed performance bottlenecks and optimized backend processing and data workflows.
    • Python
    • Backend APIs
    • OpenCV
    • Performance optimization

  3. First · Python & security

    Python & Cybersecurity Intern

    Cisco AICTE Internship to

    • Wrote Python automation scripts and data-handling utilities during a six-month technical internship.
    • Applied network security, threat analysis and secure coding fundamentals in hands-on assignments.
    • Python
    • Automation
    • Network security
    • Secure coding

Skills, grouped by what I can build with them.

No proficiency bars. Each group names the tools and links to the role or project where I used them.

See the projects View Resume (opens in a new tab)

Retrieval & semantic search

Finding the right context for a request: embedding documents, storing the vectors and searching them by meaning.

  • Embeddings
  • Vector databases
  • FAISS
  • Chroma
  • Semantic search

Used in 3EDGAR Legal Research CaseFox

Python backend & APIs

The services that make an AI workflow part of a product, and the data layer under them.

  • Python
  • FastAPI
  • Django
  • REST APIs
  • Microservices
  • SQL
  • MySQL
  • SQL query optimization

Used in CaseFox Zomit 2Database Search

Computer vision & automation

Image-processing pipelines that run in real time inside application services, and Python tooling that removes manual steps.

  • OpenCV
  • Computer vision
  • Python automation

Used in Zomit Cisco AICTE Internship

Tools & platforms

Everyday engineering tooling, plus working knowledge of two clouds.

  • Git
  • Docker
  • Postman
  • CI/CD
  • Logging
  • Monitoring
  • JavaScript
  • AWS (working knowledge)
  • Azure (working knowledge)

Used in Day-to-day tooling

Background: education and recognition

Years in software development
3+
Currently
Software Development Engineer, CaseFox
Focus
Generative AI, RAG, Python & FastAPI
Based in
Bhopal, India

  • Degree · 2021 to 2024

    B.Tech, Computer Science Engineering

    Specialization in Artificial Intelligence & Machine Learning. Lakshmi Narayan College of Technology.

  • Diploma · 2019 to 2021

    Electrical and Electronics Engineering

    Polytechnic diploma. Rabindranath Tagore University.

  • Award · 2025

    Rising Star Award

    CaseFox. Recognized for contributions to AI-powered product development.

Hiring an AI engineer, or adding AI to a product?

I am interested in AI Engineer, Applied AI Engineer and Software Development Engineer roles, and in talking with teams that are building LLM features into their products. Email is the best way to reach me.

keshavkahar7987@gmail.com

Write an email