I make generative models fast enough to ship.

01

HedgeMind

Autonomous agents running a portfolio on live market data.

Built for Inter IIT Tech Meet 14.0, on the Pathway problem statement, with the Cynaptics Club team. Kafka streams feed live prices, Reddit and Twitter sentiment, and FRED macro data into Pathway-based agents that reason over the stream and act. SARIMAX and Chronos handle forecasting, a custom MCP layer wires agents to models, and a Django + Next.js dashboard tracks P&L and anomaly alerts in real time. Not a backtest notebook: a streaming system that holds up while the market moves.

Pathway · Kafka · agents + MCP · SARIMAX / Chronos · Inter IIT 14.0

github.com/CynapticsAI/Pathway_InterIIT_14 ↗demo video ↗

02

Curio

Crawl any website and talk to it.

The language model runs entirely in your browser over WebGPU (WebLLM), so nothing you ask ever leaves your device. Not as a policy promise, as an architecture: there is no server to send it to. Embedding it on a site takes one line of code. It is also a working demo of the thing I do at fal: inference that runs where the user is.

WebGPU · WebLLM · on-device · one-line embed

curio-umber-five.vercel.app ↗

live widget · pending embed

03

LLM Distillation & Pruning

Knowledge distillation for models the size of Llama 3.1 70B, on hardware that shouldn't allow it.

Two stages. First, run the teacher once and cache its logits. Then train the student against the cached distributions on KL divergence, so the teacher never has to share the GPU with the student. VRAM stays low; performance stays put. Built at Cynaptics Club, IIT Indore.

Llama 3.1 70B scale · logit caching · KL divergence · low VRAM

04

BossForge

Photograph your messy desk. Fight it.

Your mess, your monster: an AI pipeline catalogs the clutter in the photo, designs a multi-phase boss out of it, renders it as a 3D model, and drops you into an arena against your own coffee mug. Image understanding, 3D generation, and a game stacked end to end. The demo arena has five pre-forged bosses, no upload needed.

photo → 3D boss · arena combat · 2–4 min per forge

bossforge-one.vercel.app ↗

05

Chitra

Take a selfie, get every photo of yourself out of a ten-thousand-image event gallery.

Event galleries are where photos go to disappear: thousands of images, no names, no order. Chitra searches them by face. One selfie in, every photo you appear in out. I built it alone, idea to production: the matching pipeline, the infrastructure, the product.

face matching · solo, idea → production

chitra.cloud ↗