FULL DISCLOSURE: Passion projects help me learn new technology.
I build MVPs for personal usecases and add features incrementally as per fitment.
Only a handful are deployed, with full production ready features, due to cost constraints and priority.
For example, evaluations are not in all projects. Middlewares are added only
when I needed HITL, or redact PII from RAG ingested documents, etc.
Website chatbot to answer general queries using PDF as source. 2 step RAG. Deployed on Docker, AWS EC2. Live on portfolio website via Chainlit. Incremental features added - Evaluations, Prompt Caching, Cohere Reranking for RAG documents.
RAG n8n Cohere Reranking Evaluations & Scoring LangSmith Pinecone
Read More
Inspired by a YT tutorial, this multimodal project allows you to chat with a AI persona 'Ava'. Ava has a hypothetical day schedule which is injected via context engineering to chat accordingly in different times of day. When asked to share an image of the day, a hypothetical scenario is first created and then a corresponding image is generated. Voice is given using Eleven Labs voice IDs.
AI Companion Bot Context Engineering LangGraph ElevenLabs WhatsApp Qdrant
Read More
I setup a comfy UI rig on an EC2 instance to play around with image generation using open source Flux, SDXL models and LoRAs. While setting up manually was excruciating, it was a good learning curve. Nowadays, docker images are available for easy setups.
Stable Diffusion ComfyUI LoRA AWS
Read More Blog
Most multimodal GPTs were not able to generate the new Hyundai Tucson model with the new wing-like DRLs. They generated only 2022 model. Probably trianing data cutoff. I decided to create a LoRA model that generates accurate Hyundai Tucson models.
LoRA Training Flux Replicate
Read More Blog
Ingest health params from Google Drive via dedicated ingestion pipeline. Track trends of KPI markers using middlewares. Visualise KPI trends using dashboards.
LangChain LangGraph Middlewares Ingestion-Pipelines AWS-QuickSight D3JS
Read More
Agentic AI app to search AI job openings, extract keywords and map to weekly trends. Uses HITL LangChain middlewares to confirm before saving data to Notion database.
Agentic AI Multi-Agent Notion Documents HITL Trends D3JS
Read More
Simple portfolio website. Headless CMS so it is easily configurable with new data. There are subdomains that populate subdomains such as daily news, etc.
Headless CMS Stablo Vercel
Read More
Create a digital me to manage my LinkedIn. Use my blogs and emails as dataset for SLM training using Hugging Face Transformers + TRL and PyTorch. The project increases LinkedIn engagement significantly by posting content, replying to comments and liking and sharing relevant posts. It also aims to follow topics, pages and people who have similar interests.
HuggingFace Transformers SLM TRL PyTorch LangChain
Read More
Inspired from YTs "Micheal Automates" videos. Use various LLMs to generate PineScript on TradingView to generate indiciators and signals. Then compares accuracy with historical data via backtesting. Eventually pivoted to simple AI agentic app for daily news to provide insights into swing trading equities.
AI DeepSeek Gemini TradinView YouTube PineScript
Read More
Highly SEO optimised blogs generated by AI agents using open source LLMs. Predefined blog structure was used to generate blogs. The challenge was that the blog tone was mundane and robotic. Considering SFT training SLM with actual Substack blogs, sent emails to create personality, human tone and vocabulary. Blog title is given via Slack/ Telegrat botfather, content would be saved to headless CMS Sanity that updates blog site automatically. Turns out nothing beats writing on your own on Substack.
AI Multi Agent RelevanceAI Make.com Slack
Read More
Android application to intercept phone calls from unknown numbers without ringing. Automated voive will receive call to determine if spam call via locally installed trained LLM. Conversation is transcribed on screen, in case intervention is needed. If LLM determines genuine call, will proceed to ring.
SLM Android Studio Kotlin Android-NDK Android-ADB Llama.cpp
Read More
Automated load injection, chaos injection, monitoring, reporting to give back application team members 5-6 hours everyday during game-day activities for Chase Travel.
AWS Step Functions Lambda FIS Gremlin DynamoDB Office DB-Streams
Read More
Reduced onboarding time from 4 days to less than 8hrs. Architected a multi cloud provisioning system to automate onboarding of client data to TeraData systems with modular features and validation. Reduced operational costs from $80,000 to less than $3000 every weekend. Created dashboards to show cost optimisation charts. Created scripts to auto shutdown test VMs and resources over the weekend.
Azure AWS GCP On-Demand Office Infrastructure-Provisioning
Read More
