AI / ML Engineer

Building
Intelligence
That Ships

Designing and deploying AI systems — from custom transformer architectures to production-grade multi-agent pipelines.

View Projects GitHub Profile
scroll

01   About

Bridging Research
and Production

I'm a AI/ML engineer obsessed with turning complex models into systems real users can rely on. My work spans the full spectrum, from implementing transformer architectures from first principles to orchestrating multi-agent workflows in production.

I gravitate toward local-first, privacy-respecting architectures and open-source stacks. If a problem can be solved without sending data to a third-party API, I'll find that path first.

Currently focused on LLM post-training (SFT, SDPO), RAG systems, and making state-of-the-art AI accessible to real-world applications.

Projects
11+ AI/ML Projects
A mix of LLM systems, classic ML, and full-stack applications shipped end to end.
Focus
4 Core Domains
LLMs, agents, applied ML, and production systems.
Build Style
GPT Built from Scratch
Deep interest in first-principles implementation and low-level model behavior.
Cost
0¢ Inference Cost
Strong preference for local-first systems when the problem allows it.

02   Projects

Featured Work

End-to-end AI systems: each one ships, runs, and solves a real problem.

⚖️
Kanoon Ke Haath — Indian Legal RAG Assistant
A production-hardened RAG pipeline for Indian legal Q&A, indexed against real Supreme Court of India judgments. An 11-node LangGraph orchestration (route → cache → guardrail → retrieve → rerank → generate → escalation → memory) routes questions between cheap and expensive Groq models by difficulty, behind a resilience gateway with token-bucket throttling, backoff, and circuit breaking. Guardrails cover PII, prompt injection, and citation/grounding verification; a Next.js chat UI exposes live routing, cost, and latency telemetry for every answer.
FastAPI LangGraph Redis Stack Groq Next.js Guardrails & Evals
11
LangGraph Pipeline Nodes
4
Guardrail Layers
2x
Cost-Aware Model Routing
100s
SC Judgments Indexed
📄
FinRAG — Financial Document Intelligence System
Upload a financial PDF, extract page-aware text, chunk it for retrieval, embed it locally, and store it in Qdrant. Retrieves the top-15 chunks, reranks to top-3 with Cohere, and answers with exact page citations using Groq. Containerized for AWS EC2, traced end-to-end with LangSmith.
FastAPI PyMuPDF Qdrant Cohere Rerank RAGAS v0.4
⚗️
SDPO LLM Fine-Tuning Pipeline
3-checkpoint post-training pipeline (Base → SFT → SDPO) on Llama-3-8B. SDPO uses the model's own verified outputs as a dense training signal — no external reward model or teacher required. Trained with Unsloth 4-bit QLoRA, cutting VRAM usage ~30% via gradient checkpointing.
SDPO Unsloth QLoRA TRL PEFT LLM Post-Training
🧠
Custom GPT-2 from Scratch
Full GPT-2 implementation in PyTorch — multi-head attention, layer norm, GELU, residual connections — with support for loading official 124M pre-trained weights and top-k sampled text generation.
PyTorch Transformers tiktoken NLP
🤖
Automated Insurance Data Entry
Multi-agent RAG pipeline that reads PDFs, extracts structured data, and stores it — all with local Llama 3.2 inference. Zero API costs, fully private, production-ready.
CrewAI LangChain ChromaDB Ollama RAG
🎙️
Interview Enhancer
AI-powered mock interview tool with real-time NLP analysis and feedback. Evaluates responses for clarity, relevance, and confidence — 4 ML pipelines including CNN facial analysis and LSTM voice profiling at 87% accuracy.
Python CNN LSTM Multi-modal
📈
ARIMA Stock Market Predictor
Time-series forecasting on 5,033 days of AAPL data (1998–2018). Implements AR, MA, ARMA, and ARIMA with Dickey-Fuller stationarity testing, log-scaling, differencing, and inverse transformations.
pandas statsmodels scikit-learn Time Series
Also shipped
Mistral-7B Fine-Tuning Pipeline
Tumour Classifier (CNN)
Employee Turnover Prediction
Auto-Forecaster Pipeline
Traffic Sign Classification
Gallery-Vite (TypeScript)

03   Skills

Technical Stack

The tools I reach for — and know well enough to build around.

LLM Training & Alignment
Unsloth / QLoRA PEFT / LoRA TRL SFT SDPO DPO Quantization (4-bit / GGUF) Hugging Face Hub
Deep Learning & NLP
PyTorch Transformers GPT Architectures Multi-Head Attention Text Generation Embeddings tiktoken
AI Systems & Agents
CrewAI LangChain Multi-Agent RAG Pipelines ChromaDB Ollama Llama 3.2
ML & Data Science
scikit-learn statsmodels ARIMA pandas numpy matplotlib CNN
Full-Stack Dev
Python Flask PostgreSQL SQLAlchemy TypeScript Vite Vanilla JS

04   Decisions

Why the system is built this way

P.01
Local embeddings
all-MiniLM-L6-v2 keeps embeddings cheap, fast, and private while still providing strong semantic search quality.
P.02
Two-stage retrieval
Vector search maximizes recall, then reranking filters for precision so the generator sees better evidence.
P.03
Structured generation
JSON mode and Pydantic guarantee a stable answer schema with explicit page references instead of freeform text.
P.04
Traceability first
LangSmith traces and RAGAS scores make the tradeoffs visible, so quality can be tuned with evidence.

05   Contact

Let's build something worth shipping

Open to interesting problems in AI/ML, full-stack, or anywhere they intersect.

GitHub Email