ankit-keshari

I buildAI systemsthat ship.

Agents, ML pipelines, and local-first tools that actually run.

AI Intern @ EY · B.Tech + AI/ML minor, NIT Delhi '27

See projects
Get in touch

Open to internships & freelance AI work

Ankit Keshari, laughing, looking away from the camera
portrait.distort()still

projects

Things I've shipped

Agents, ML systems, and tools I actually run. The numbers on these cards are measured, not claimed.

SUNDAY — Job-Hunt Agent — An agent that hunts jobs while I sleep.

SUNDAY — Job-Hunt Agent

An agent that hunts jobs while I sleep.

Every morning at 10:00 IST it scrapes AI/ML intern and fresher openings, ranks them against my real resume, builds ATS-safe resume and per-company cover-letter PDFs, and emails me the top five. I reply with a company name — it applies. It never emails the same company twice.

Runs live for me daily — the deployment stays private (it holds my resume data).

  • Vercel Serverless
  • Node.js
  • Supabase
  • Gmail SMTP/IMAP
  • cron
GitHub ↗

scroll · drag · ← → to explore

experience

Where I've worked

Two internships, two very different sides of applied ML — attacking models, then building with them.

  1. AI Intern · EY

    Jun 2026 — Present

    Building an LLM-powered document-intelligence platform for a government client — end to end, from scanned PDFs to defensible scores.

    • OCR pipeline for scanned official journals: PyMuPDF page rendering + vision-model transcription, cached so no document is read twice.
    • Rubric-driven LLM scoring that quotes its evidence — every score carries the exact line in the source document it came from.
    • Evaluated against 136 expert-rated products: exact-match and within-one accuracy tracked per parameter to pick the best model.
    • FastAPI
    • PostgreSQL + pgvector
    • LlamaIndex
    • Gemini Vision
    • Next.js

  2. Research Intern — Adversarial Machine Learning · DRDO — Scientific Analysis Group (SAG), Delhi

    May — Jul 2025

    Attacked face-recognition models to measure how they break — black-box adversarial ML at a defence research lab.

    • Engineered black-box adversarial attack pipelines (SimBA, NES) across 8+ models on 9,000+ samples.
    • Implemented transfer-based PGD attacks with norm constraints, using cosine similarity for evaluation.
    • Designed score-based query attacks that iterate on similarity feedback alone.
    • Evaluated model robustness with DeepFace embeddings and analyzed attack transferability across architectures.
    • Python
    • PyTorch
    • DeepFace
    • NumPy

next

Building next

Fully planned, honestly labelled — these are blueprints, not vaporware demos.

Designed in full · build starts next

CodeLens

See your code think.

Write DSA code and watch it execute live — memory, stack frames, heap objects, pointer arrows, recursion trees — animating as you type. One language-agnostic trace engine drives every panel; Python, C++, and Java plug into it. Reverse execution comes free: the whole run is a scrubbable timeline.

  • React
  • TypeScript
  • Monaco
  • Pyodide
  • WASM

Planned · PTB-XL · Kaggle GPUs

ECG Interpreter

Deep learning on 12-lead ECGs — explained honestly.

A 1D-CNN/ResNet trained on PTB-XL to flag cardiac abnormality classes, paired with deterministic signal features (intervals, rhythm, axis) and a grounded explanation layer that can only say what the classifier and measurements support. Educational tool — never a medical device, and it says so on every screen.

  • PyTorch
  • wfdb
  • neurokit2
  • RAG

about

Behind the terminal

I'm Ankit — a mechanical engineering student at NIT Delhi who fell down the AI rabbit hole and never climbed back out. Most days I'm building agents and ML systems: a job-hunt agent that applies while I sleep, a fully offline dictation app, trading bots that grade their own predictions against the market.

I care about two things: tools that actually ship, and numbers that are actually real. Every project here runs on measured results — hit rates, latencies, accuracy against expert baselines — never vibes.

Off-screen: badminton, lawn tennis, volleyball, and finding new places to travel.

education
B.Tech Mechanical Engineering, Minor in AI & ML — NIT Delhi, 2023–2027
responsibility
Database Sector Head, Training & Placement Cell, NIT Delhi (Coordinator 2023–25)
currently
DSA in Java, daily — plus an Obsidian second brain that runs my life
Ankit laughing with a hand over his eyes
candid.exe — no retakesstill

skills

Tools of the trade

No percentage bars — either I've shipped with it or it's not listed.

languages

  • Python
  • Java
  • C
  • MATLAB

ml & data

  • PyTorch
  • TensorFlow
  • Scikit-learn
  • DeepFace
  • NumPy
  • Pandas

core

  • Adversarial ML
  • Regression Modeling
  • Data Structures & Algorithms

shipping with

  • Next.js
  • TypeScript
  • Supabase
  • FastAPI
  • Cloudflare Workers
  • Ollama
  • faster-whisper
  • Three.js

tools

  • Claude Code
  • Jupyter
  • Figma
  • Canva

contact

Let's build something

Recruiting, freelance AI work, or just an idea worth building — my inbox is open.

ankitkeshari326@gmail.comGitHubLinkedInResume ↗