AJ Aditya Jethani
AI/ML Engineer

I ship intelligent
systems to production.

I am Aditya Jethani, an AI/ML engineer with 1.5+ years turning ambiguous problems into deployed RAG, agentic, computer-vision and backend systems across legal-tech, healthcare and e-commerce. I optimise until latency, accuracy and robustness targets are actually met.

View work Resume
epoch by epoch
Portrait of Aditya Jethani
Surat, India · PDEU '25
CGPA 9.44 / 10 · GATE CS 95.61%ile
9×RAG inference speedup
139.6 tok/ssustained generation
360Kdocs in agentic RAG
94.2%tongue-CV accuracy
0.99token-F1, doc parsing
2071LeetCode Knight rating
9×RAG inference speedup
139.6 tok/ssustained generation
360Kdocs in agentic RAG
94.2%tongue-CV accuracy
0.99token-F1, doc parsing
2071LeetCode Knight rating
About

I build the unglamorous parts that make AI products work in the real world: batched GPU inference, retrieval that stays faithful, evaluation that catches failures before users do.

0
years in production AI
0
shipped projects
0
B.Tech CGPA / 10
0
merged OSS PRs

I am a Computer Engineering graduate from Pandit Deendayal Energy University (CGPA 9.44/10), now an AI/ML Engineer at Logicwind. My approach is iterative and measurement-first: break a vague brief into experiments, automate the evaluation, and keep going epoch by epoch until the numbers hold.

I work end to end, from data pipelines and model fine-tuning (QLoRA) to FastAPI services, agent orchestration and the dashboards that prove a system is behaving. I care most about why systems fail, and I build the tooling to find out.

  • [ rag ]

    Retrieval & Agentic Systems

    Faithful RAG, chunking and embedding evaluation, multi-agent orchestration with observability.

  • [ cv ]

    Computer Vision

    Medical imaging diagnostics, OCR pipelines, detection and tamper analysis.

  • [ infra ]

    ML Systems & Delivery

    GPU-batched inference, FP16 quantization, FastAPI/Go services, cloud deployment.

Experience

Where I have shipped.

Four roles across three companies, each ending in something deployed or measured.

Dec 2025 — Present
Surat, India
Full-time

Machine Learning Engineer · Logicwind

  • Built a production multi-agent system on the OpenAI Agents SDK with LangFuse for observability, tracing and prompt versioning across concurrent workflows, cutting manual prompt-tuning time 60% and reaching 90% response satisfaction.
  • Architected an agentic RAG platform over 360K multilingual documents (45% scanned, Surya OCR), reaching 4.3× throughput via async batching and Qwen3 embeddings; benchmarked 16+ chunking strategies and 6+ embedding models.
  • Engineered an agentic NL-to-SQL voice interface that returns governed SQL plus auto-generated chart dashboards in under 2s, giving non-technical stakeholders self-serve analytics.
  • Drove applied research on deterministic document parsing that beat VLM-based parsers on native PDFs (CER 0.0135, token-F1 0.99); manuscript in preparation.
  • Shipped production AI across OpenAI Agents SDK, Google ADK, Mastra AI and Vercel AI SDK with Docker deployment, cutting client agent-delivery time 70%.
OpenAI Agents SDKGoogle ADKLangFuseQwen3RAGFastAPIDocker
Jun 2025 — Dec 2025
Delhi, India
Full-time

Software Engineer, AI/ML · CybraneX

  • Engineered a high-throughput RAG inference pipeline for QLoRA fine-tuned LLMs, achieving a 9× speedup (latency 110s → 12.2s) by replacing serialized per-thread generation with GPU-batched inference.
  • Sustained 139.6 tokens/sec and 0.82 queries/sec in production through FP16 quantization, KV-cache, FlashAttention-2 and context truncation.
  • Built a clinical tongue-analysis computer-vision system (ResNet18 + custom coating model, OpenCV, Fourier and Gabor analysis) reaching 94.2% color and 91.7% coating accuracy, served via a REST API with an interactive UI for clinical workflows.
  • Delivered an enterprise document-extraction product deployed on Azure, combining Cloud Vision OCR, tamper detection and LLM classification across 14+ document types with a CI/CD pipeline.
  • Added model-safety and evaluation tooling that surfaced 27% more quality violations via statistical drift detection and structured observability.
PyTorchQLoRAvLLMOpenCVFastAPIAzureCI/CD
Jan 2025 — Jun 2025
Surat, India
Internship

Software Engineering Intern · Yanolja Cloud Solutions

  • Architected an image-to-image translation system for multilingual e-commerce that lifted accuracy 50% over the CycleGAN baseline while preserving native typography and layout.
  • Built an automated code-review system (privately hosted on Google Cloud for GitHub PRs) that mined commit histories and PR diffs, mitigating 27% of recurring security vulnerabilities through RAG-based pattern detection.
  • Engineered Go + Python + FFmpeg computer-vision pipelines for large image datasets, reducing average processing time 1.3×.
GoPythonFFmpegLangChainRAGGCP
May 2024 — Oct 2024
Delhi, India
Internship

Machine Learning Intern · CybraneX

  • Engineered an AI telemarketing system processing 50,000+ daily calls with custom LLMs, reaching 89% accuracy in legitimate-lead identification under sub-200ms latency.
  • Built a real-time analytics dashboard blending purchase history and geolocation signals, raising a client's inventory turnover 22%.
  • Accelerated healthcare document extraction from 5s to 0.5s per document through batch optimization and pipeline consolidation.
LLMsPythonAnalyticsETL
Selected work

Things I built, with the numbers.

A curated set spanning agents, multimodal CV, RAG and scientific ML. Every figure below is from the real project.

PRJ-01 TSllama-3.3-70b · groq

TalentScout AI ↗

Production candidate-screening assistant: parses resumes, generates role-specific questions and scores answers with running confidence.

85% validation accuracy-60% hiring time
Python · LangChain · LangGraph · Llama 3.3 70B · Streamlit
Prompt Detective interface PRJ-02

Prompt Detective ↗

Multimodal pipeline that reverse-engineers the prompt behind an AI-generated image or video using CV signal extraction and semantic retrieval.

85% Top-K accuracy<5s / 1-min video
OpenCV · CLIP · Sentence-Transformers · pgvector
Grade Flow interface PRJ-03

Grade Flow ↗

RAG teacher's assistant that automates grading with plagiarism detection, auto question-paper generation and a timed test environment.

90% workflow automated-75% eval time
FastAPI · LangGraph · pgvector · FAISS · Gemma-2
PRJ-04 CVresnet18 · coatingnet

Tongue Analysis AI ↗

Clinical computer-vision system for traditional-medicine diagnostics: coating, color and crack analysis over 2,524 images, replacing subjective scoring.

94.2% color acc.91.7% coating acc.
PyTorch · OpenCV · scikit-image · Flask
Scrollwise interface PRJ-05

Scrollwise ↗

Full-stack digital-wellbeing app and Chrome extension that turns scroll, click and focus events into real-world analogies with AI insights.

<500ms insight latencyMV3 extension
Next.js · Express · TypeScript · MongoDB · Groq
PRJ-06 ODEmimic-iii · torchdiffeq

Neural ODE · Heart Rate ↗

Continuous-time ODE model reconstructing and forecasting irregularly sampled ICU heart-rate trajectories from MIMIC-III with masked-point imputation.

MSE 0.443RMSE ≈ 7.01 bpm
PyTorch · torchdiffeq · MIMIC-III
See all projects
Research & writing

Papers and applied research.

I am drawn to evaluation, faithfulness and failure analysis of LLMs in retrieval-augmented and multilingual settings.

Deterministic Document Parsing that Outperforms VLM Parsers

A zero-ML, failure-mode-driven PDF parser reaching CER 0.0135 and token-F1 0.99 on native PDFs, with a self-improving optimizer that lowered character-error rate from 17.07% to 8.29%. Applied research at current work.

In preparation

Advanced Forecasting of Solar Power Generation using Bi-LSTM with Fourier Features

Bi-LSTM with Fourier and geo-temporal features for daily AC power forecasting, achieving R² = 0.86 and RMSE = 128.09 kW.

Under review

Efficient Labelling of Hinglish Datasets for Hate-Speech Classification

A consensus labeling pipeline (RoBERTa, mBERT, VADER) with active learning and RLHF feedback for low-resource code-mixed text, reaching 86% test accuracy.

Manuscript in prep
Toolkit

The stack I reach for.

Languages

  • Python
  • Go
  • C++
  • SQL
  • JavaScript
  • TypeScript
  • Bash

ML / AI

  • PyTorch
  • TensorFlow
  • scikit-learn
  • QLoRA Fine-tuning
  • Agentic RAG
  • LangChain
  • LangGraph
  • Transformers
  • OpenCV
  • vLLM

Agents & LLM Infra

  • OpenAI Agents SDK
  • Google ADK
  • Mastra AI
  • Vercel AI SDK
  • LangFuse
  • MCP

Cloud & Data

  • GCP (Vertex AI, BigQuery, Dataflow)
  • AWS (EC2, S3, EMR)
  • Azure Web Apps
  • Docker
  • ETL / ELT
  • CI/CD

Backend & APIs

  • FastAPI
  • Flask
  • Django
  • Node.js
  • Express
  • Next.js
  • Pydantic

Databases & Vector

  • PostgreSQL
  • MongoDB
  • MySQL
  • pgvector
  • FAISS
  • ChromaDB
  • Qdrant
Proof

Competitions, contests and open source.

Knight
LeetCode
Max rating 2071 (top 1.75%), 190+ solved, AIR 61 in Weekly Contest 462 among 35K+.
Top 105
Google Solution Challenge 2025
Among 4,000+ international teams worldwide.
95.61%ile
GATE CS 2026
Graduate Aptitude Test in Engineering, Computer Science.
Top 10
Hackathon · IIT Patna
Top-10 finish at ByteVerse 2025.
Top 10
Hackathon · DAIICT
Top-10 finish among a large field of teams.
Finalist
JPMorgan Code for Good 2024
Plus Smart India Hackathon 2023, Samsung Solve for Tomorrow, Flipkart GRID 5.0 and DotSlash 6.0 / 7.0.
Rank 936
CodeChef Starters 113D
Top 5% among 19K participants. Kaggle rank 278, Season 4 Ep. 8.
15+ PRs
Open Source
Merged in Hacktoberfest 2024 and GSSoC 2023 / 2024.
Credentials

Certifications.

Five Anthropic programs on Claude and agents, deep-learning fundamentals, and presented summer-school research.

A\
Claude 101
Anthropic
A\
Claude with the Anthropic API
Anthropic
A\
Introduction to Model Context Protocol
Anthropic
A\
Introduction to Agent Skills
Anthropic
A\
AI Fluency: Framework & Foundations
Anthropic
NV
Fundamentals of Deep Learning
NVIDIA DLI
CIS
IEEE CIS Summer School
Presented · IIT Indore
ACM
ACM Summer School
IIT Gandhinagar · NIT Goa · IIT Patna
{ }
ML with Python
freeCodeCamp
G
Google Skills Profile
view ↗
Community

Mentoring and leadership.

2023 — 2025
ACM · PDEU Student Chapter

Advisor & Chair

Mentored 50+ developers and ran automation workshops, DSA bootcamps and code-a-thons. Hosted hands-on labs on Selenium, RegEx, LangChain, Git and systems design.

2021 — 2024
GeeksforGeeks · PDEU Chapter

Tech Lead

Produced hands-on labs on Git, GitHub and shell scripting for 150+ learners, and ran competitive-programming intensives and interview-prep sessions.

2021 — 2023
Cube-i-Cult

Documentation Head

Published event recaps and SOPs and ran logistics for national and inter-college cube-solving competitions.

Writing & video

I explain what I learn.

Breakdowns on Medium and YouTube, from activation functions to neural architecture search.

logits & reason · 01
✎ Medium

Logits & Reason: Part 1

Mar 2026
logits & reason · 02
✎ Medium

Logits & Reason: Part 2

May 2026
meta llm compiler
✎ Medium

Meta LLM Compiler, simplified

Jul 2024
aligngpt · multimodal
✎ Medium

Understanding AlignGPT: gateway to multimodal intelligence

Jun 2024
✎ Medium

Activation Functions: coffee for neurons

Mar 2024
✎ Medium

Neural Architecture Search, explained

Jan 2024
▶ YouTube

How to score 97/100 on your resume: tips & tricks

Apr 2024
▶ YouTube

AQI data forecasting using LLMs

Mar 2024
▶ YouTube

Medicinal plant identification system

Oct 2023
Read more on Medium
Contact

Always keen for a
good collaboration.

✦ adityajethani11@gmail.com in / LinkedIn ⌥ GitHub ◆ Kaggle ✎ Medium ▶ YouTube