Skip to content
Saad Khazi Ahmed
Résumé(PDF download)

AI & backend engineering · Hyderabad

Saad Khazi Ahmed

I build AI systems from the data up, and the software that puts them to work.

I spent the summer of 2025 on the data pipelines behind a Telugu large language model. My own projects go from a model that reads handwritten maths into LaTeX to a web app that tests and deploys itself to Kubernetes on every push.

Selected workGet in touchGitHubLinkedIn
Now
Final year, B.Tech CSE (AI & ML) at VJIT, Hyderabad
Previously
AI Developer Intern, VISWAM.AI — Telugu LLM initiative
Works across
NLP data, ML models, LLM tooling, backend, CI/CD
Open to
Internships and early roles in AI and backend engineering

01About

I'm studying Computer Science with a specialisation in AI and machine learning at Vidya Jyothi Institute of Technology in Hyderabad, and graduate in 2027.

Most of what I know I learned by taking things all the way through: training a model and then working out how to evaluate it, wrapping an LLM in a tool someone could actually use, putting a small web app behind tests, containers and a Kubernetes deployment.

In 2025 I interned at VISWAM.AI on the Telugu LLM initiative with Meta, IIIT Hyderabad and Swecha Telangana. My work was the data side of the model — cleaning, deduplicating and validating speech and text corpora before any of it reached training.

I want to work on products that people use, somewhere I can move between models, backend services and the infrastructure that runs them.

Education

  • Vidya Jyothi Institute of Technology2023 — 2027

    B.Tech, Computer Science & Engineering (AI & ML)

    Hyderabad · CGPA 8.96 / 10

  • Narayana Junior College2021 — 2023

    Intermediate, MPC

    Kurnool · 93%

02Experience

May — Jun 2025

Hyderabad

AI Developer Intern, VISWAM.AI

Telugu LLM initiative — Meta × IIIT Hyderabad × Swecha Telangana

  • Built NLP data pipelines for a national-scale Telugu large language model, working across multilingual speech and text datasets.
  • Wrote the preprocessing workflows that cleaned, annotated and deduplicated raw data and assembled it into training corpora.
  • Designed ingestion pipelines to bring new multilingual datasets into the corpus at scale.
  • Worked with AI researchers on dataset validation, model evaluation and the data engineering in between.
Where the work sat
  1. 01Speech + text sources
  2. 02Clean
  3. 03Annotate
  4. 04Deduplicate
  5. 05Validate
  6. 06Training corpus

03Selected work

Now building · in progress

Open source · building it with @syedmukheethRepositoryfor SIEdits on GitHub (opens in a new tab)

SIEdits

An open-source video editor that your AI assistant can operate

Timeline editing and motion graphics in the browser, for talking-head creators. Every cut, caption, zoom and graphic is a typed, validated operation, so the creator's own assistant (Claude, ChatGPT, Cursor) can edit through an MCP connector, and every change can still be undone or tweaked by hand.

I'm building this with @syedmukheeth. My work so far: fixing the Windows setup, hardening the local API, and stripping generic UI patterns out of the editor so it reads like professional editing software.

TypeScript / React 19 / Hono / Zod / MCP / WebCodecs / Whisper (WebGPU)

Week 1 — go / no-go spikes before product code

  1. A

    Can the browser decode footage, draw captions and export a 1080×1920 MP4?

    Valid H.264 + AAC output. Under 1× real time once, 2.4–2.7× under sustained load.

    Works, needs speed

  2. B

    Are Whisper (WebGPU) word timestamps good enough for word-level captions?

    31% within ±100 ms raw. A constant +145 ms lag; 89–92% after correcting it.

    Needs correction

  3. C

    Can an AI client drive edits through a remote MCP server with typed ops?

    Full agent edit round-trip passes locally. OAuth and deployment still open.

    Passes locally

First results on one laptop, as recorded in the repository.

A

Handwriting to LaTeX

Handwritten Mathematical Expression Recognition

Academic project
Handwritten imagePreprocessVisual encoderAttention decoder

Decoded token by token

  1. \frac
  2. {
  3. x
  4. ^
  5. {
  6. 2
  7. }
  8. +
  9. 1
  10. }
  11. {
  12. \sqrt
  13. {
  14. y
  15. }
  16. }

Typeset result

x2 + 1√y
Illustrative input → output. Not a recorded model run.

Problem

Handwritten maths is two-dimensional. Fractions, exponents and radicals stack and nest, so line-by-line OCR can read the symbols but loses the structure that gives them meaning.

What I built

A recognition system that takes an image of a handwritten expression and produces structured LaTeX that can be typeset, searched or evaluated.

  • Image preprocessing to normalise handwriting before it reaches the model.
  • OCR combined with an attention-based sequence model that emits the expression as LaTeX tokens.
  • Training and evaluation tuned across multiple handwritten datasets.

Why sequence modelling

Predicting a LaTeX token sequence, with attention deciding where to look at each step, lets the model recover nesting and position together instead of classifying symbols in isolation.

Python / Deep learning / Computer vision / OCR

B

Cold Email Agent

LLM-powered outreach writer

Personal projectRepositoryfor Cold Email Agent on GitHub (opens in a new tab)
Sender
Who are you? e.g. a SaaS copywriter
Offer
What do you offer?
Target
Who are you targeting?
Tone
friendly · casual · corporate
Mistral 7B
via OpenRouter

Subject

Bodyunder 120 words

CTA

Follow-upnext day

Inputs and output format, as defined in the code.

Problem

Personalised outreach takes time to write well, and the generic templates that save that time are the ones people ignore.

What I built

A Python agent that turns four inputs (who is sending, what they offer, who they are writing to, and the tone) into a complete message: subject line, a body under 120 words, a call to action, and a follow-up for the next day.

  • OpenAI-compatible client routed through OpenRouter to Mistral 7B Instruct.
  • A library of 25 industry-specific prompt templates.
  • The wider workflow connects generation to REST APIs, Supabase and Vercel for storage and delivery.

Fixed output contract

The prompt pins the response to the same four parts every time, so output can be parsed and sent without hand-editing. Going through an OpenAI-compatible endpoint means the model can be swapped without touching the code.

Python / LLMs / Prompt engineering / REST APIs / Supabase

C

Smart Parking System

Real-time slot monitoring on NodeMCU

Hardware project
Slot sensorNodeMCULive slot status
Occupied Free
Schematic of the system, not live data.

Problem

In most car parks nobody knows which slots are free until they drive past them, so drivers circle and queues build at the entrance.

What I built

An IoT system that watches 10+ parking slots and keeps their status up to date automatically, so availability can be read at a glance.

  • A sensor at each slot, read by a NodeMCU board.
  • Embedded firmware turns raw readings into occupied / free states.
  • Live status updates without anyone checking slots by hand.

Hardware meets software

Most of the work sat at the boundary: getting dependable detection out of physical sensors and turning it into a status people can trust.

IoT / NodeMCU / Embedded systems / Sensors

D

Commit to Kubernetes

Containerised web app with automated CI/CD

Personal projectRepositoryfor Commit to Kubernetes on GitHub (opens in a new tab)
  1. Push
  2. Test
  3. Build image
  4. Smoke test
  5. Push to registry
  6. Rolling deploy

GitHub Actions

Deploys to a temporary kind cluster inside the runner, then checks /health.

Jenkins

Same stages, defined as code, deploying to the local Kubernetes cluster.

Pods scale 2 → 6 at 70% CPU · rollback on failed rollout

Pipeline stages, from the repository's workflow files.

Problem

Deploying by hand is how untested builds reach users and how a bad release turns into downtime.

What I built

A Node.js web app that is tested, containerised and deployed to Kubernetes automatically on every push, through two independent pipelines: GitHub Actions and Jenkins.

  • Jest tests with coverage, then a multi-stage, non-root Docker image with a health check.
  • Each image is smoke-tested before it is pushed, and tagged by commit SHA so it can be traced.
  • Rolling updates with zero unavailable pods, readiness and liveness probes, and an automatic rollback if a rollout fails.
  • Horizontal Pod Autoscaler (2–6 pods at 70% CPU). Jenkins is configured entirely as code.

Testing the deploy, not just the build

Cloud runners can't reach a cluster on a laptop, so the Actions pipeline spins up a throwaway kind cluster inside the runner, deploys the new image and checks /health. Jenkins deploys the same image to the real local cluster.

Node.js / Docker / Kubernetes / Jenkins / GitHub Actions

Also on GitHub

04Skills

Languages
  • Python
  • Java
  • Go
  • C
  • SQL
  • JavaScript
AI / ML
  • Machine learning
  • NLP
  • Large language models
  • Prompt engineering
  • Computer vision
  • OCR
Backend
  • Flask
  • Django
  • REST APIs
  • Supabase
Cloud & DevOps
  • AWS
  • Docker
  • Kubernetes
  • Jenkins
  • Vercel
  • Git & GitHub
Data
  • Power BI
  • Excel
Foundations
  • Data structures & algorithms
  • OOP
  • DBMS
  • Operating systems
  • Computer networks

05Recognition

Hackathons

  • FinalistHackForge 2.0ISL Engineering College
  • ParticipantCodeStreet HackathonAmerican Express
  • ParticipantBharatiya Antariksh HackathonISRO

Certifications

  • Cloud Engineering, Data & Analytics (CEDA) LaunchpadPwC
  • Machine Learning using PythonInfosys
  • Red Hat Linux FundamentalsRed Hat
  • Python EssentialsCisco
  • C EssentialsCisco
  • Data Analytics EssentialsCisco

06Contact

Hiring, building something, or curious about one of these projects? Email is the quickest way to reach me.

saadkhaziahmed2005@gmail.com
LinkedInGitHubBased in Hyderabad, India · IST (UTC+5:30)