I Build ML Systems That Are
Boring To Maintain And
Interesting To Read.
AI/ML Intern @ PGAGI // Building production-style backend systems and AI-driven services -- REST APIs, relational/vector databases, and multi-agent LLM orchestration // Open to Backend & AI-GenAI Roles
> Loading tensors... OK.
01 // SYSTEM_OUTLOOK
"I am deeply obsessed with the gap between a machine learning model that scores 94% on an offline benchmark and one that actually survives production deployment. I don't build generic tutorial wrappers. I design clean pipelines, implement architectures from scratch to truly understand their boundaries, and build systems optimized for reliable execution over academic metric inflation."
02 // FIELD_DEPLOYMENT
DOWNLOAD RESUME ↓AI/ML Intern
@ PGAGI- -> Built and optimized a WhatsApp-based AI document processing pipeline processing 10K+ invoices/month with 98% extraction accuracy, reducing manual validation time by 65%.
- -> Improved document-processing reliability by implementing media validation, PDF-to-image conversion, per-user batching, and file-size safeguards, supporting robust processing of image and multi-page PDF inputs.
- -> Evaluated LLM/VLM performance across invoice, image, and multi-page PDF scenarios by measuring extraction accuracy, latency, token usage, and inference cost, informing model and pipeline optimization.
- -> Contributed to backend, frontend, and production engineering through REST APIs, asynchronous processing, database migration planning, React functionality, testing, logging, and debugging, while researching Excel-to-Python business-logic conversion for product integration.
03 // SYSTEM_CAPABILITIES
Languages
Python, C++, JavaScript, SQL.
Backend & APIs
FastAPI, Flask, REST APIs, Async Processing, Microservices, Git/GitHub, Testing & Debugging.
Databases
PostgreSQL, MySQL, MongoDB, SQLAlchemy, Alembic, pgvector, FAISS.
AI/GenAI & Agentic Systems
PyTorch, HuggingFace Transformers, Sentence-Transformers, RAG, Agentic AI, LangChain, LangGraph, LLM/VLM Integration.
ML/DL & Transformers
Scikit-learn, TensorFlow, CNNs, Transformers, Attention, Encoder-Decoder, BERT, GPT, ViT, Model Tuning.
Cloud & Tools
Google Cloud (GCS, Vertex AI), Docker, Streamlit, NumPy, Pandas, OpenCV, AWS Certified Cloud Practitioner.
04 // PROJECTS
Vendor Onboarding & Compliance Microservice
Built a FastAPI microservice with 6 REST endpoints for document ingestion, LLM extraction, compliance validation, duplicate detection, pipeline orchestration, and audit-log retrieval. Combined Groq (Llama 3.3 70B), Sentence-Transformers, FAISS, and PostgreSQL with Alembic to automate structured extraction, rule-based compliance checks, semantic duplicate detection, and audit logging.
IMPACT:
Reduced manual processing time by 60%, automated compliance validation for 500+ vendors.
PodFlow — Multi-Agent Newsletter Pipeline
A 4-stage multi-agent pipeline orchestrating transcript extraction, insight generation, newsletter drafting, and email delivery, converting YouTube videos into reviewable newsletter drafts. Added 2 human-in-the-loop checkpoints for content review and send confirmation, reducing reliance on single-call LLM generation.
IMPACT:
Cuts newsletter creation from 2 hours to 15 minutes, enabling batch processing of 50+ videos/month.
Semantic Search Engine with Neural Reranking
An end-to-end semantic retrieval pipeline combining Sentence-Transformers with pgvector/PostgreSQL for dense vector search without relying on LangChain abstractions. Implemented a 2-stage retrieve-then-rerank architecture using bi-encoder retrieval and cross-encoder reranking, deployed as an interactive Streamlit application to improve search relevance.
IMPACT:
Achieves 98% search relevance vs. baseline 60%, sub-100ms latency at scale.
GPT-2 Style LLM — Built from First Principles
Implemented a complete GPT-2 style transformer in PyTorch from scratch, building BPE tokenization, multi-head self-attention, feed-forward layers, positional encodings, and layer normalization. Built pretraining and fine-tuning pipelines with gradient clipping, learning-rate scheduling, mixed-precision training, and checkpointing, then loaded pretrained GPT-2 weights to validate implementation correctness.
IMPACT:
Validated implementation by loading OpenAI weights, achieving 100% architecture parity with original.
PodFlow (Restructuring)
I am entirely restructuring PodFlow into an AI-powered knowledge workspace that turns heterogeneous resources into grounded, structured knowledge artifacts that users can review, edit, share, export, send, and store.
CV Optimization (CIFAR-10)
Highly optimized CNN focused on memory efficiency and parameter reduction via pruning techniques for edge deployment.
Real-time Facial Detection
Real-time facial detection pipeline optimized for low latency inference on live video streams.
Autoencoder from Scratch
Neural autoencoder built from the ground up exploring representation learning, latent space compression, and data reconstruction.
Research Logs &
Architectural Deep Dives
How Transformers Actually Work: From Search Bar to ChatGPT
A complete architectural breakdown — from the query-key-value attention mechanism to how modern LLMs like GPT and BERT are trained and fine-tuned.