hi, I'm

Ankit Raj Mehta

AI Developer · final-year Computer Engineering student · Lalitpur, Nepal

I build AI systems, mostly around LLMs and agents, and I'm just as drawn to the core of it: machine learning, deep learning and the maths underneath. Most recently I was a Junior AI Engineer at iBriz.ai, building AI agents and async AI services with FastAPI, Celery and MongoDB.

Mostly, though, I'm curious. I'll happily lose an evening to history, physics, a maths puzzle or a documentary about octopuses, and I like taking things apart to see how they work: a vision-language model rewritten in plain PyTorch, a memory library for LLM apps, little creatures that evolve their own brains. I'm in my final year at IOE Pulchowk Campus.

download CV ↓

linkedin github email

Projects

things I've built, with the people I built them with

demo coming soon
memBlocks
A modular memory library for LLM apps. Memory lives in attachable blocks, each with core facts, hybrid-searched semantic memory and a rolling summary, plus an MCP server and a LoCoMo eval harness.
- Python
- LLM memory
- Qdrant
- MCP
- FastAPI
› notes
Three memory sections per block: core memory (persona and user facts, always injected), semantic memory (extracted facts, searched with dense + SPLADE vectors in Qdrant and fused with RRF), and a recursively compressed conversation summary.

Background updates: when the message window fills, core extraction, semantic extraction with ADD/UPDATE/DELETE/NOOP conflict resolution, and summary regeneration run in parallel.

Ways in: a Python library, a FastAPI + React demo app, and an MCP server so assistants such as Claude Desktop can use a block.

Eval: a LoCoMo benchmark harness. The upstream README reports 1,027 memory tokens per query, against 1,764 for Mem0 and 26,031 for full context.
video sooncode ↗ for memBlocks on GitHub team repo ↗ for memBlocks
demo coming soon
PaliGemma from scratch
Google's PaliGemma vision-language model rebuilt from scratch in PyTorch, a SigLIP vision encoder joined to a Gemma decoder, then loaded with the official pretrained weights.
- PyTorch
- VLM
- Transformers
- SigLIP
- Gemma
› notes
Vision side: a SigLIP ViT encoder (modeling_siglip.py) that turns an image into patch embeddings.

Language side: the Gemma decoder with a KV cache, plus the linear projector that merges image tokens into the text sequence (modeling_gemma.py).

Real weights: a loader maps the Hugging Face paligemma-3b-pt-224 safetensors onto the hand-written modules, so the model answers prompts about an image.
video sooncode ↗ for PaliGemma from scratch on GitHub
demo coming soon
Desktop Automation Agent
A desktop assistant you drive by voice or chat. A LangGraph ReAct agent picks from system, Gmail, web-search and browser-use tools to act on the machine.
- LangGraph
- Agents
- FastAPI
- browser-use
- Playwright
› notes
Agent: a LangGraph ReAct agent with memory, served by a FastAPI backend to a web chat UI. A separate voice loop takes spoken commands.

Tools: system stats, volume, screenshots, opening and closing apps, clipboard and keyboard shortcuts, a calculator and Tavily web search.

Sub-agents: a Gmail agent that drafts and sends mail, and a browser-use agent that drives a real Chromium through Playwright.
video sooncode ↗ for Desktop Automation Agent on GitHub team repo ↗ for Desktop Automation Agent
demo coming soon
ProtoEvolve
A C++/SDL2 neuroevolution sandbox. Creatures with small evolving neural networks learn simple tasks over generations, with speciation that protects new network wirings.
- C++
- SDL2
- Neuroevolution
- Genetic algorithms
› notes
Brains: each creature carries a feed-forward network whose weights and active connections both mutate.

Tasks: a timed right-then-left migration, or chasing a food box that keeps moving. The mode is a template argument.

Speciation: genomes with sufficiently different wiring start a new species, which is protected from extinction for a while so new ideas get time to improve.

Rendering: SDL2 replays creature trails every few generations and the best genome at the end.
video sooncode ↗ for ProtoEvolve on GitHub
demo coming soon
MindTune Brainwave
EEG data collection with a Bluetooth headset, and a calm / neutral / stressed classifier (LightGBM, ExtraTrees) running in a live desktop UI.
- EEG
- LightGBM
- scikit-learn
- Python
- Tkinter
› notes
Collection: reads the headset’s ThinkGear stream over Bluetooth serial, and a Tkinter collector lets an annotator tag emotions and events with hotkeys while recording.

Features: band-power percentages, band ratios (theta/beta, alpha/beta, slow/fast) and rolling statistics.

Live app: a Tkinter window reads the headset, runs the saved model and shows the predicted state. It lives on the classifier branch.
video sooncode ↗ for MindTune Brainwave on GitHub team repo ↗ for MindTune Brainwave
demo coming soon
Sign-Ease
A hackathon app that turns hand signs into text, from MediaPipe landmarks and a TFLite model behind a React + Flask video app.
- MediaPipe
- TFLite
- React
- Flask
- Hackathon
› notes
Pipeline: the React client records short clips, and the Flask API extracts MediaPipe landmarks and runs a TFLite sign-recognition model on them.

Context: built in a 36-hour hackathon. It won Overall Winner at Sagarmatha HackFest and was presented at LOCUS 2024.
video sooncode ↗ for Sign-Ease on GitHub team repo ↗ for Sign-Ease

Experience & education

iBriz.ai
San Francisco crypto/AI company
- Junior AI EngineerFeb 2026 – Apr 2026
- AI InternAug 2025 – Feb 2026
- Built AI agents, including an LLM-driven GEO (Generative Engine Optimization) analysis agent.
- Built async AI systems with FastAPI, Celery and MongoDB.
- Worked on the Omorenda project, experimenting with Openclaw-based agents for longevity and wellbeing.
eRuchi Pvt. Ltd.
Kathmandu market-research startup
- Technical Operations and Data AssociateJan 2025 – Jul 2025
- Data Analytics and AI Research InternSep 2024 – Jan 2025
- Built a content-based recommendation engine.
- Led a team of 3 developers on a data-collection platform.
- Did data analysis and visualisation.
PDSC
Part-time, software development
- Innovation Team MemberApr 2023 – May 2024
Navya Nepal
Volunteering
- Founding MemberMay 2020 – Feb 2023
- Ran donation and student-support programmes.

education

Tribhuvan University, IOE Pulchowk Campus
- BE in Computer Engineering2023 – 2027
St. Xavier's School, Jawalakhel
- +22020 – 2022

Odds & ends

quick facts

BE Computer Engineering, IOE Pulchowk (2023 – 2027)
Fusemachines AI Fellowship, Microdegree in AI
3× hackathon grand winner

toolbox

AI/ML
Python
RAG
TensorFlow
FastAPI
Celery
Git
REST APIs
SQL
Pandas
Matplotlib

hackathons

Grand Winner, Yathartha Dataverse · Thapathali Campus, with Aayush Mishra
Grand Winner, Sagarmatha HackFest · Sagarmatha Engineering College, with SignEase
Grand Winner (FinTech), LOCUS Hack-A-Week 2024 · Pulchowk Campus, team Looptroop
Presented SignEase at LOCUS 2024

certificates

Fusemachines AI Fellowship, Microdegree in AI · Apr 2025 – Dec 2025
Machine Learning Specialization · Stanford & DeepLearning.AI
Supervised Machine Learning · Coursera, Oct 2023
Advanced Learning Algorithms · Coursera, Nov 2023
Unsupervised Learning, Recommenders, Reinforcement Learning · Coursera, Nov 2023
MLOps Concepts · DataCamp, Sep 2024
CS50 · Harvard
Intro to TensorFlow for AI, ML and DL · DeepLearning.AI

teaching & fellowships

Fusemachines AI Fellowship (Apr – Dec 2025) · a hybrid, 20-week fellowship in AI taught by Fusemachines
Taught a machine-learning session at the LOCUS 2026 Software Fellowship
Selected for the Data Fellowship 2024 (Cohort III) · Sunway College Kathmandu, with DataCamp Donates

The one-page version

My CV as a PDF, last updated July 2026.

cv.pdf ↓

Things I've written

27 Oct 2026
Rebuilding PaliGemma from scratch
I rebuilt Google's PaliGemma (SigLIP vision encoder + Gemma decoder) in PyTorch, and the most interesting part was the attention mask.
10 Jun 2026
MemBlocks: memory for LLMs that you control
A modular memory architecture for LLMs, with domain-scoped blocks, agentic memory management and an MCP server.
13 Jun 2025
How we proved eRuchi's recommendation engine works, with real data
A controlled experiment comparing personalised recommendations against random suggestions at eRuchi.

all posts →

Say hi

happy to talk AI/ML work

The quickest way to reach me is a message on LinkedIn. Email and GitHub work too.

Ankit Raj Mehta

Projects

memBlocks

PaliGemma from scratch

Desktop Automation Agent

ProtoEvolve

MindTune Brainwave

Sign-Ease

Experience & education

Experience

iBriz.ai

eRuchi Pvt. Ltd.

PDSC

Navya Nepal