AI / ML & Backend Systems Engineer
Yadidiah
Kanaparthi
I build production-shaped AI and backend systems — RAG pipelines, LLM fine-tuning, voice agents, and the FastAPI/Postgres services that hold them up. I work end to end, with tests, migrations, and eval harnesses rather than notebooks, and I go deep on how the models actually behave.
- 1+ yr
- AI / ML engineering
- 6
- Systems built end to end
- 3
- Peer-reviewed papers
- 16
- Certifications
Now
Fine-tuning Qwen2.5-1.5B for schema-conditioned text-to-SQL with an execution-accuracy eval harness.
Selected work
Six systems built end to end — data pipeline to interface. Each links to a spec sheet: the problem, the approach, and what shipped.
- 01
GraphRAG Hierarchical Chat
Graph-structured retrieval over long, cross-referenced document sets
PythonLangGraphNeo4j / graph storeEmbeddings2026RAG / Retrievalprototype - 02
PneumoScan AI
Explainable pneumonia detection from chest X-rays
TensorFlowKerasGrad-CAMStreamlit2024Computer Vision / Medicalshipped - 03
YouSentimentAI
End-to-end MLOps pipeline for YouTube comment sentiment
LightGBMTF-IDFMLflowDVC2024MLOpsshipped - 04
LexiQE AI
Private, local legal-document Q&A
RAGFAISSFLAN-T5Sentence-Transformers2024RAG / NLPshipped - 05
Decoder-only GPT, from scratch
A small transformer built and trained in raw PyTorch to learn the internals
PyTorchNumPyTransformers (theory)2025Deep Learning / Fundamentalsresearch - 06
Premium Business Planner
AI-assisted, investor-ready business plans with live analytics
Gemini APIStreamlitPlotlyPandas2025Generative AI / Productshipped
Stack
Grouped by what I actually reach for, not a logo wall. Depth is uneven and honest — strongest on backend and RAG, learning-by-building on training internals.
LLM / GenAI
- RAG (FAISS · pgvector · Pinecone)
- LoRA / QLoRA — Unsloth
- LangChain · LangGraph · DSPy
- Prompt contracts + eval harnesses
- Hugging Face · Ollama
- PyTorch (from-scratch transformers)
Backend
- Python · FastAPI
- PostgreSQL · SQLAlchemy · Alembic
- Redis · RQ workers
- Outbox pattern · idempotency keys
- NestJS · TypeScript
- Structured logging · readiness checks
ML / Data
- TensorFlow · Keras · scikit-learn
- XGBoost · LightGBM
- MLflow · DVC
- Pandas · NumPy · Apache Spark
- Power BI · Tableau
Voice AI
- LiveKit · Pipecat
- Deepgram · Whisper
- Silero VAD
- Dual-STT validation
Cloud / Ops
- GCP — Cloud Run · GKE · Vertex AI
- AWS — SageMaker · Lambda
- Docker · Kubernetes
- GitHub Actions CI
Experience
One full-time role, plus job simulations from the pre-graduation stretch. Kept separate so it's clear which is which.
Sep 2025 — Present
Humai
AI Automation Engineer
- Build multi-step agent workflows with LangChain and LangGraph — tool use and control flow as an agent loop, not a linear chain — for document processing and validation tasks.
- Fine-tuned a 5B Gemma model (QLoRA via Unsloth) on a 13K-example bilingual domain dataset: chat-template formatting, response-only loss masking, inference validation — owned from data prep to deployment.
- Built a real-time voice agent on Pipecat (Deepgram STT, GPT-4, Silero VAD) with a parallel Whisper-validation layer that cross-checks transcription without adding response latency.
- Designed a 6-stage compatibility-matching pipeline (FastAPI + MongoDB Atlas Vector Search, Sentence-BERT and CLIP scoring, auction-based assignment) — 10K profiles in ~8 minutes at a 70–80% match rate.
- Built an MCP gateway so external clients — Claude, Cursor, ChatGPT, custom agents — call internal tools through one interface.
- Ship the FastAPI services and infra around all of it: LiveKit interview agents, a CV-analysis pipeline scoring 200+ candidates, deployed on cloud container platforms.
Job simulations · Forage
2025
PwC · Forage
Digital Intelligence (job simulation)
- Built Python classification models and analysed feature importance.
- Produced valuation and cash-flow forecasting documents.
2025
Deloitte · Forage
Data Analytics (job simulation)
- Designed interactive Tableau dashboards.
- Classified data in Excel to extract business insights.
2025
British Airways · Forage
Data Science (job simulation)
- Analysed customer review data for drivers of purchasing behaviour.
- Built a predictive model on the review dataset.
2025
BCG · Forage
GenAI Consulting (job simulation)
- Built an AI financial chatbot over 10-K / 10-Q filings.
- Rule-based logic layered on parsed financial data.
2024
Quantium · Forage
Data Analytics (job simulation)
- Processed transaction datasets for commercial insight.
- Identified benchmark stores for uplift testing.
Open source
Pulled live from the GitHub REST API and cached for an hour — public repos, language mix, and what was pushed most recently.
33
public repos
3
stars earned
9
followers
Python
top language
Language mix · by repo count
- Python 33%
- Jupyter Notebook 17%
- JavaScript 17%
- HTML 17%
- TypeScript 6%
- Svelte 6%
Recently pushed
- LLM-ProjectsPython · ★ 1 · 4mo ago
- Youtube-Sentiment-AnalysisPython · ★ 1 · 1y ago
- DataScience-analysis-miningJupyter Notebook · ★ 1 · 1y ago
- Yadidiah-k— · ★ 0 · 23d agoConfig files for my GitHub profile.
- graphrag-hierarchical-chatPython · ★ 0 · 29d ago
- finance-trackerTypeScript · ★ 0 · 4mo ago
Writing & notes
Long-form posts on shipping AI systems, plus a running set of technical notes on how LLMs, transformers, and retrieval actually behave — written from first principles and kept accurate against implementation.
Notes · on this site
- N1
Deep learning, stated precisely
The parts that are easy to get almost-right: what a neuron computes, the difference between backprop and gradient descent, and why stacked layers build a feature hierarchy nobody designed.
Neural networks · Training · Fundamentals — 6 min
- N2
The transformer forward pass, token by token
From raw text to a contextual representation: tokenization, embedding lookup, Q/K/V, attention, and why stacked layers refine rather than rewrite.
Transformers · Attention · Embeddings — 8 min
- N3
What actually happens at inference
The generation loop, why decoding parameters are post-hoc, how the KV cache works, what GQA and quantization trade away, and how one model serves many requests at once.
Inference · KV cache · Serving — 10 min
- N4
Building RAG from the boundaries in
Retrieval and grounded generation are two steps, not one. Define the interfaces first — chunking, embeddings, vector store, retriever, generator — and treat every model output as untrusted input.
RAG · Retrieval · AI systems — 9 min
- N5
Backend patterns for AI work
AI calls are slow, expensive, and unreliable. The backend around them needs durable jobs, worker-owned state, retry-safe writes, and a way to recover when a queue drops a message.
Backend · Durable jobs · Reliability — 8 min
Published · Medium
Jun 2025
Building production-ready LLM applications with RAG
What it actually takes to move a RAG demo into production: chunking that respects structure, retrieval you can inspect, and citation validation before anything reaches the user.
Apr 2025
MLOps best practices: from Jupyter to production
Turning experimental models into systems with MLflow, DVC, and CI — so every result is reproducible from a commit hash.
Feb 2025
Deep learning for medical imaging: explainable pneumonia detection
Building a CNN for chest X-rays where the Grad-CAM overlay matters as much as the accuracy score.
Research
Three peer-reviewed papers from my undergraduate work, on machine learning for IoT security and log-based intrusion detection.
- 2024IEEE
Machine-learning approaches for security threat detection in IoT networks
IoT security · anomaly detection
- 2024Springer
Log analysis for intrusion detection using ML classifiers
Log analysis · classification
- 2023IEEE
IoT–ML integration for smart monitoring systems
IoT · sensor data · edge
Credentials
Degree, then the certifications worth listing. The full set is longer; these are the ones that map to real work.
Aug 2021 — May 2025
B.Sc. Computer Science — Magna Cum Laude
Canadian University Dubai
- Specialised in AI and Machine Learning.
- Coursework: Deep Learning, Computer Vision, NLP, MLOps.
- 3 peer-reviewed papers published (IEEE, Springer).
- All-semester Dean's List. Led a 200+ member ML club.
2020 — 2021
A / AS-Level — Computer Science, Mathematics, Physics
GEMS Winchester School, Dubai
- School topper in Computer Science.
- Mathematics Olympiad winner.
2018 — 2020
IGCSE — ICT, Mathematics, Physics
GEMS Winchester School, Dubai
- Practical lab work across ICT, Physics, and Chemistry.
Certifications · 16 listed
- Prompt Design in Vertex AIGoogle CloudJun 2025
- Certified Software EngineerHackerRankApr 2025
- Generative AIMicrosoft · LinkedInMar 2025
- Data AnalysisMicrosoft · LinkedInMar 2025
- Data Science Professional CertificateAnacondaMar 2025
- Data Science Professional CertificateKNIMEMar 2025
- Intro to Deep LearningKaggleNov 2024
- RESTful API (Intermediate)HackerRankDec 2024
- Software Engineering Job SimulationAccenture Nordics · ForageMar 2025
- Data Science Job SimulationBritish Airways · ForageMar 2025
- Data Analytics Job SimulationDeloitte Australia · ForageMar 2025
- Digital Intelligence Job SimulationPwC Switzerland · ForageMar 2025
- Data Analysis with PythonfreeCodeCampJun 2022
- Scientific Computing with PythonfreeCodeCampMay 2022
- Artificial Intelligence & Machine LearningAWSNov 2022
- MATLAB OnrampMathWorksMar 2023
Contact
Open to AI engineering & RAG/agent work. The fastest way to reach me is this form or LinkedIn.
Elsewhere
Dubai, AE · 25.20° N, 55.27° E