AI Engineer · Gujarat, India ·
I make language models work in production. Fine-tuning, quantization, serving, evaluation, and agents: at Nextbase Solutions I build GenAI products that carry real revenue, and on my own time I publish experiments with the evals to back them.
live: this page is training a small AI on my writing, right now, on your device. scroll down to watch it learn, or press ` to talk to it.
epoch 00 · watch it learn
It learns while you read
When this page opened, a tiny neural network woke up on your device knowing nothing. It has been teaching itself to write like me ever since. Scroll to replay everything it has learned.
the model is trained. now meet it.
epoch 01 · selected work
Selected work
innerlens
Anthropic released Jacobian-lens interpretability research; within a day I shipped the production runtime on PyPI. It reads a model's own activations while it generates and flags fabrications the text alone cannot reveal.
hallucination signal at AUROC 0.80 02MathTutor-Qwen3-8B
Fine-tuned Qwen3-8B into a K-12 math tutor, then built the LLM-as-judge pipeline that grades it. Along the way I caught the judge rewarding the wrong behavior and fixed the methodology.
beat the base model on 5 of 5 metrics 03HGD Memory Engine evaluation
Benchmarked an agent memory engine through six versions and contributed the fix that closed the last gap. Recall went from 40% to 85% while context tokens dropped by 78-93%.
recall 40% to 85% across 6 versions 04Rhizome Logic
A competitive-intelligence agent that runs itself: scheduled collection, per-fact source citations, confidence scores, and a human gate that keeps hallucinated intel out of the record.
3 scheduled agents, about $23/monthAvatar Restaurant Ordering
Dine-in customers scan a QR code and order by talking to an animated avatar. Multi-tenant SaaS, now live.
NestJS · Next.js · liveAI Bill Splitter
Gemini vision parses receipt photos into line items. You assign them by talking: "Bob and Charlie shared the pizza."
React · Gemini · open sourceAI Storybook Generator
Gemini image pipeline that keeps characters consistent across scenes, with resumable checkpoints and backoff under rate limits.
Node.js · BullMQ · private buildepoch 02 · experience
Experience
AI Engineer · Nextbase Solutions
LLM serving with 4-bit AWQ on vLLM, measured at 2-3x faster generation on a 24 GB GPU. Diffusion media tools (Flux, Wan 2.2, ComfyUI on serverless GPUs) that account for over 30% of company revenue. A multi-provider LLM gateway with 50-74% context compression. AI agents for three departments.
Software Engineer, CV & Robotics · Techno Smart Diamond Solution
Vision control core for an autonomous diamond-polishing robot: mesh registration, per-facet pose math, and corrections streamed to the robot's PLC through a pybind11 C++/Python bridge.
Software Engineering Intern · Techno Smart Diamond Solution
3D reconstruction of rough diamonds from camera and laser scans, down to 5-micron error. Facet-line detection with DBSCAN clustering and RANSAC regression.
Research Intern · Uka Tarsadia University with CIT Kokrajhar
Blockchain messaging DApp for educational institutions with IPFS file sharing and end-to-end encryption.
epoch 03 · skills
What I work with
- Fine-tuning & serving
- QLoRA, PEFT, AWQ quantization, vLLM, Hugging Face, NEFTune [02]
- Interpretability
- Jacobian-lens workspace readouts, internal-confidence scoring, hallucination signals [01]
- Evaluation
- LLM-as-judge pipelines, ablations, benchmark design, judge-bias auditing [02] [03]
- Agents & retrieval
- LangChain, LangGraph, LlamaIndex, CrewAI, RAG, Qdrant, FAISS, Chroma [04] [03]
- Media generation
- ComfyUI, Flux, SDXL, Wan 2.1/2.2, RealESRGAN, serverless GPU workers
- Engineering
- Python, PyTorch, scikit-learn, TypeScript, Node.js, Next.js, FastAPI, Docker
- Cloud
- AWS, GCP, Vertex AI, Azure, RunPod, Firebase, Supabase
converged · contact