Aakash Madabhushi
I build applications on top of language models, with retrieval that grounds them and gates that stop them acting alone.
Retrieval-grounded systems are the core of it: a Pinecone index of CVE and MITRE ATT&CK material feeding Llama 3.3 to explain why a network link was flagged, returning an analyst-readable answer in under a second; and a scheduled pipeline that embeds articles so they can be searched by meaning rather than keyword.
The agent work takes a consistent stance, which is that a person approves before anything irreversible happens. InvoSense scores its own confidence on each extracted field and blocks every payment action pending human review. The LinkedIn recruiting assistant takes a recruiter from résumé to outreach but pauses before it sends. Reader runs its model on-device, so nothing you read leaves your machine. I have also worked a layer below the framework, having written a GPT-2 and a byte-pair tokenizer from scratch, sped up attention, and fine-tuned Phi-3 into a chat assistant. I ran a LangChain workshop series for SJSU’s AI/ML Club.
Every link goes to code or data you can inspect yourself.
Blocks 100% of payment actions until a person approves them, by reading invoice amounts and dates with Llama 3.2 Vision, scoring its own confidence in each field, and gating every write behind a LangGraph human-approval step.
Personal project · Jun 2026
Detects an API failure, diagnoses a root cause, and acts on it without a human, by wiring FastAPI health probes into a Railtracks LLM agent that tunes its own alert thresholds from prior incidents. Built in five hours at a hackathon.
SJSU Applied Data Science Hackathon 2026 · Mar 2026
Kept résumé parsing and embedding matching off the request path across 17 containers, by routing them through Kafka and caching profile reads in Redis for 6× faster repeat lookups, with an AI assistant that pauses for human approval.
SJSU DATA 236, Distributed Systems · Spring 2026
Keeps everything you read on your own machine, with zero external API calls, by running Gemini Nano on-device inside a Chrome MV3 extension paired with a SwiftUI iOS app that tracks focus sessions across both.
Personal project · Jul 2026 to present
Reached 0.90 F1 flagging suspicious host pairs in 27.1 million Zeek flows, by condensing them into a connection graph with Airflow, dbt and Snowflake, classifying links with a graph neural network, and having Llama 3.3 explain each alert.
SJSU capstone, team project · Feb to May 2026
Makes articles searchable by meaning rather than exact keywords, by embedding them with Sentence Transformers on an Airflow schedule and serving nearest-neighbour lookups from a Pinecone index.
Personal project · Nov 2025
Select a project for the full detail and links.
I'm available from December 2026 and open to relocating. Email me and I'll reply within a day.
Hiring for something else?