Built, evaluated, and handed over so your team can run it without me.
beta · you're chatting with Karan's AI twin — it can be wrong, so double-check anything that matters
Scroll down ↓
RAG system
Ask about SEC filings in plain language. Every answer cites the passage it came from, and it tells you when the filings can't answer the question.
Retrieval pipeline
Python · FastAPI · LangGraph · PydanticAI · pgvector · RAGAS · Gemini
Agentic SaaS
An agent for owner's reps running construction bids. It reads the RFP, pulls scattered contractor questions into one table, and recommends a vendor.
Bid pipeline
Next.js · FastAPI · LangGraph · Gemini · pdfplumber · ReportLab
A wider sample of the work, built alongside the RAG projects and all public: an agent skill, clinical NER, EEG classification, generative image models, and time-series RUL.
Projects
A Claude Agent Skill that runs a startup idea through five stages and stops at the first one it fails. Each stage scores out of 10, nothing under 7 passes, and about a quarter of ideas clear it. One verdict, a full case report, no coaching.
Fine-tunes BioBERT for token-level clinical NER on the MACCROBAT2018 case reports: about 40 entity types across 78 BIO tags, scored on entity-level F1. Runs on Apple Silicon (MPS) out of the box.
Detects Alzheimer's, FTD, and MCI from EEG two ways: time-frequency spectrograms classified by attention CNNs (CBAM, SE, self-attention), and PSD features run through Random Forest and XGBoost. Saliency maps and SHAP explain both. A methods comparison, not a benchmark.
Convolutional VAEs on three datasets. MNIST for generative modeling with a walkable 2D latent space; COCO and CelebA for polygon-mask inpainting, scored with PSNR and SSIM.
Remaining-useful-life on NASA IMS bearings and PHM 2010 tool wear, always tested on a machine the model never saw in training. Traditional ML (Random Forest, XGBoost) and deep nets (CNN-LSTM, TCN, wavelet-CNN) compared under that leave-one-out constraint.
Predicts the remaining life of an AK-47 recoil spring from gunshot audio: as the spring fatigues, the gap between muzzle blast and bolt closure stretches. A digital-twin simulator turns a handful of seed recordings into thousands of run-to-failure trajectories, then XGBoost, a 2D CNN on scalograms, and a 1D CNN on raw audio each predict RUL.
Stack
I’m Karan, a RAG and agentic AI engineer. I help teams get unstuck from their own documents. I work calm and methodical. I’d rather spend a few days finding the right approach than rush a wrong one into production. And I don’t think AI replaces people. It’s more like a sharp kid that takes you word for word: genuinely useful when you’re clear with it, a liability when you’re not. So I keep a human in the loop and hand over systems a team can actually run. Off the clock, it’s the gym and coffee, in that order.
Useful information gets stuck, and once you notice it you see it everywhere. It’s buried in a policy document no one can find, or it lives only in the head of the one person who actually knows how things work. When they’re out, everyone waits. When they leave, the knowledge goes with them. I grew up around this in India, where a lot still runs on paper and manual sign-offs, but it shows up everywhere, whether that’s a courtroom or a corner shop. That’s what I build against: RAG systems that pull the right answer out of the mess and cite where it came from, and agents that take the slow, manual document work off people’s plates.
Based in Pune, India· Available for freelance & contract work
Testimonials
I’ve enjoyed working with Karan for 1.5 years, and his dedication and positive attitude amaze me.
Background
EarlierApril 2026 – Now
Building RAG systems, AI agents, and AI workflows for clients. I take each one from scoping through evals to a handover the team can run without me.
Independent · Pune
Jan 2025 – Present
Going under the hood of the models: architecture, how they actually work, and how to modify them. Deep learning and data science, studied while I ship.
M.S. Ramaiah Univ. · GPA 9.6
July 2022 – Jan 2025
Shipped RAG that held up with real users: hybrid search plus a query-rephrasing step that fixed multi-turn context loss.
Software Developer · Amdocs · Pune
Jan 2022 – June 2022
Made streaming sensor data trustworthy: telemetry from 1,000+ industrial sensors, ingested and monitored live.
Project Intern · Sisai · Pune
Aug 2018 – July 2022
Led the electronics team at Vegapod, a student hyperloop project, on control systems.
B.Tech ECE · MIT-WPU · Pune
Available for freelance & contract
Tell me what’s slowing your team down, and we’ll figure out where AI can actually help.
Résumé (PDF) ↗