Neel Ghoshal

A router that uses Jev, TypeSafe AI's decision model, to decide when a cheap small LLM (qwen3:1.7b) is good enough and when a prompt needs a larger one (qwen3:8b), plus an evaluation against simple baselines on a cost versus accuracy curve. It compares two designs on 400 prompts from MMLU, MMLU-Pro, ARC-Challenge and GSM8K, with thresholds tuned on a dev split and 95% bootstrap intervals on the test split: pre-routing, where Jev predicts from the prompt alone whether the small model will be right, and a cascade, where Jev checks the small model's answer and escalates only the likely wrong ones. Pre-routing did no better than random routing, but Jev proved a reliable answer checker (AUROC 0.927), and the cascade cut compute time by 41% against always using the large model, at 73% accuracy against 80%.

A GPT-style decoder-only transformer written and trained from scratch on a single laptop GPU (RTX 4060, 8 GB), with no pretrained weights or pretrained tokenizer. It trains its own byte-level BPE tokenizer with 8,192 tokens on the TinyStories dataset of about 2.1 million short stories, then trains a 29.6M-parameter model (8 layers, 8 heads, 512-token context, pre-norm blocks and weight tying) with bf16, AdamW and a cosine learning-rate schedule tied to a wall-clock budget. After 6 hours and 1.44 billion tokens it reaches a validation loss of 1.285 (perplexity 3.6) and writes coherent short stories.

A monitoring and efficiency layer for how LLM agents call the tools of three biomedical MCP servers (Medical MCP, BioMCP and OMOP MCP). It records exact token usage per tool call and the pathway the model takes to reach each tool, then tests strategies that make those calls cheaper, such as schema compression, result shaping, a turn limit with a forced final answer, and an "answer early" instruction. Strategies are compared in fractional factorial experiments on a 54-task benchmark, and every accepted change carries a trust score, evidence and a versioned decision log. Over two rounds of changes, the accepted configuration roughly halved tokens per successful task (from about 52,300 to 25,000), while task success rose from 88% to 96%.

A four-agent retrieval-augmented tutor for a Mathematics for Machine Learning course, answering only from the course's lecture slides and textbook: a Router classifies each question as lecture, textbook, or off-topic; a Tutor retrieves the relevant chunks and answers with page citations; a Verifier checks every claim against the retrieved text and triggers a retry if one is unsupported; and a Quizzer writes and grades quiz questions, logging missed topics to revisit first next time. Runs entirely on a local, open-weight LLM with no API key.

A local MCP (Model Context Protocol) server that gives Claude a searchable library of research papers: add papers by arXiv ID or PDF path, and Claude can search them, answer questions with section and page citations, compare papers, and follow citation links between papers in the library. Everything runs on-device, combining vector embeddings with BM25 keyword search over a local SQLite index, with no external service beyond arXiv downloads.

A Salesforce Agentforce member service agent for Cedarline Health Plan, a hypothetical medical insurer: verifies a member's identity, then answers questions about claims, denial reasons, deductible and out-of-pocket balances, prior authorizations, service requests, and appeals, using only that member's own records. A data pipeline turns ~1,160 Synthea sample patients into Cedarline's data, adding it to Snowflake and to Salesforce Data Cloud, and feeds an Agentforce Data Library the agent searches to answer questions.

A document integrity review tool built for the ATO Hackathon: paste in a bill, resolution, or article, and four AI agents run in parallel, checking its completeness and sourcing, verifying its factual claims against live web search, confirming the authority of the people named in it, and flagging internal or external contradictions by severity. A LangGraph graph orchestrates the agents and merges their findings into one report.

A Django backend that prioritizes AI-flagged imaging findings (X-ray, CT, MRI) for radiologist review: critical, high-confidence findings surface first in the worklist. Radiologist verdicts are recorded and rolled up into a confirmation-rate stat, a rough real-world measure of how much the AI's flags can be trusted.

A machine learning web app that predicts developer salary from country, education level, and years of experience, trained on the Stack Overflow Developer Survey and served through an interactive Streamlit interface.

An implementation of "A Neural Algorithm of Artistic Style": blends the content of one image with the visual style of another using a pretrained VGG network and gradient-based optimization in TensorFlow.

A complete accessibility tool that lets visually impaired users control video calls without a keyboard or mouse: an OpenCV real-time hand-gesture pipeline maps gestures to meeting controls (mute, hand-raise) and pairs with speech recognition for audio feedback.

An assistive web platform that helps people with visual, auditory or vocal disabilities navigate and communicate in hospitals without needing a personal attendant: TensorFlow sign-language detection from a webcam feed, in-browser speech-to-text, and door detection with OpenCV image processing, with Django handling user accounts.

An online exercise-training platform built on PoseNet and p5.js. Real-time webcam posture detection tracks the user's skeleton and starts an on-screen timer only when the correct exercise posture is held, enabling accurate, self-guided personal training.

An augmented-reality furniture-shopping app built in Unity: point your phone's camera at a room or object to measure it in AR, then browse matching furniture sized to fit from e-commerce listings, or drop a 3D model of a piece into your room to preview it in real time, with metric readouts handy for a trip to IKEA.