Case studies, with the work attached

Problem

People run their lives across a dozen disconnected apps with no single signal of how any life axis is actually trending. There's no operating system for the whole person.

Build

AllOS: a config-driven life OS. LifeOS is the free entry product — 8-axis life radar, daily questions, weekly AI portrait — and modules (FounderOS, StudentOS, SalesOS, RantOS) unlock from profile signals. Turborepo monorepo: Next.js web + Expo mobile + role-gated admin, Prisma, and a Claude AI layer (classifier, scorer, connectors). 16 PRDs locked before a line of product code.

Problem

Bank loan DSAs in India depend on a freelance CAD draftsman to produce the 2D floor plan every home-loan file requires — a 1-day, ₹-per-plan bottleneck on every application.

Build

AI-assisted SaaS that turns property details or an uploaded sketch into a bank-ready 2D floor plan PDF in under 90 seconds. Credit-based pricing, multi-floor support. Next.js 14 + Clerk + Prisma/Supabase, GPT-4o + DALL·E 3, inline-React SVG plans, html2canvas + jsPDF export.

Problem

Banks and NBFCs publish repossessed-vehicle auctions across a dozen portals with no consistent format, no search, and no alerts. Buyers miss deals they'd happily pay for.

Build

One searchable platform aggregating repossessed-vehicle auctions across India. Python + Playwright scrapers normalise and dedupe listings into a tsvector-searchable Postgres schema; a FastAPI backend runs a 15-minute alert engine; Next.js 14 frontend with tier-gated contact details and a dealer dashboard.

Problem

Single-LLM calls can't decide when to compute, search, or retrieve — routing logic has to live somewhere explicit and testable.

Build

LangGraph StateGraph with typed state: a math tool-use loop plus a multi-agent router that dispatches between web search, RAG, and direct LLM answers based on query type.

Problem

Competitive research is repetitive and stateless — every session starts from zero, losing everything learned before.

Build

LangGraph + Claude Sonnet agent with MemorySaver checkpointing: retains competitor context across sessions for retail store analysis, so each run builds on the last.

Problem

Agent frameworks ship many collaboration patterns — knowing which fits which job requires building each one.

Build

Four AutoGen patterns implemented end-to-end: reflection loop for self-critique, FunctionTool BMI calculator, StateFlow portfolio manager, and a multi-modal bill analyzer.

Problem

Useful crews need tools beyond chat: live search for current data, code execution for verifiable answers.

Build

Two advanced CrewAI systems: a learning-path generator wired to SerperDev search plus a custom tool, and a code debugger using CodeInterpreterTool to actually run the code it fixes.

Problem

Logistics optimization needs two distinct kinds of reasoning — analysis and strategy — that pollute each other in one prompt.

Build

Sequential CrewAI pipeline: a Logistics Analyst agent handles VRP routing analysis, then hands off to an Optimization Strategist that produces an ABC/XYZ inventory roadmap.

Problem

Frameworks hide the reasoning loop. To trust agents in production you need to have built the loop bare at least once.

Build

Custom ReAct loop with zero framework code: Groq LLM plans the steps, Tavily executes searches, and the LLM synthesizes a final research report — plan/act/observe written by hand.

Problem

Travel answers go stale instantly — an assistant is only useful if it pulls live data and structures it.

Build

LangChain AgentExecutor combining live weather (WeatherAPI) and DuckDuckGo attraction search into a structured travel summary with consistent output formatting.

Problem

Foundational AI papers (Attention, GPT-4, InstructGPT, Gemini, Mistral) are dense and scattered — answering precise questions across them takes hours of manual cross-reading.

Build

RAG chatbot grounded in the seminal AI papers: retrieval across multiple papers with cited answers. Capstone project for the Analytics Vidhya GenAI Pinnacle program.

Problem

Most RAG tutorials hide everything behind hosted APIs — costs, latency, and zero understanding of the retrieval mechanics.

Build

Full RAG pipeline with no API keys: FAISS vector store, MiniLM embeddings, MMR retrieval for diversity, and Flan-T5-Large generation — every component local and inspectable.

Problem

Real documents break naive RAG: a 10-K spanning 2014–2023 needs reranking and question decomposition, not just top-k retrieval.

Build

LlamaIndex pipelines on LIC policy docs (3 embedding models × 2 LLMs compared) and Coca-Cola 10-K filings with a Cohere reranker and SubQuestion query engine.

Problem

Full fine-tuning of even small transformers is out of reach on consumer hardware — parameter-efficient methods are the practical path.

Build

QLoRA applied to BERT: 4-bit NF4 quantization with LoRA adapters, training only ~0.1% of the 110M parameters while preserving task performance.

Problem

Tokenization quietly shapes everything an LLM does — cost, context limits, multilingual behavior — yet most builders never look inside one.

Build

Byte-pair encoding tokenizer trained on WikiText-2: 30K vocabulary, special tokens, TemplateProcessing, packaged as a HuggingFace PreTrainedTokenizerFast.

Problem

Extracting structured data from financial statements is high-stakes — prompt technique choice materially changes accuracy.

Build

Five techniques benchmarked on ABN AMRO financials via Claude Vision: zero-shot, chain-of-thought, role-based, JSON-constrained, and multi-image prompting.

Problem

Students summarize PDFs and hand-write practice questions separately — two slow steps that belong in one pipeline.

Build

LangChain SequentialChain that ingests a PDF, produces an LLM summary, then generates an MCQ quiz from that summary — one chained flow, consistent output.

Problem

Industry analysis ($5.1B→$47B AI agent market) is read-once content — generating it as a formatted document should be automated end-to-end.

Build

Multi-LLM RAG pipeline that researches and writes an AI-agent industry report, rendered to .docx using only the Python standard library — no Office dependencies.

Problem

Model choice should be evidence-based: how much does each architectural step actually buy on a real 1.6M-row dataset?

Build

Sentiment classification ladder on 1.6M tweets: TF-IDF baseline, then PyTorch RNN, LSTM, GRU, and BERT — accuracy and cost compared at each rung.

Problem

Water quality monitoring needs both a continuous index and a class label — two models double the maintenance burden.

Build

Single dual-task deep network on CPCB water data: WQI regression and quality classification from one shared representation, built in TensorFlow/Keras.