AI/ML EngineerResearcherOpen-source systems
AI/ML Engineer and researcher working across deep learning, LLM systems, agentic AI and machine-learning research.
01Selected work
Open-source runtimes, frameworks and evaluation harnesses for LLM applications, presented with source-linked evidence, documentation and stated limitations.
Cognitive runtime for AI agents
Published on PyPI · runtime in alphaA neuroscience-inspired cognitive runtime that gives long-lived agents typed memory, budget-aware retrieval, working memory, consolidation and safety behind one small Python API.
50-turn prompt-token benchmark
Source (opens in a new tab)424,361 → 79,381
Author-reported simulated conversation; counted with tiktoken (cl100k_base).
Grounded LLM annotation
Live demo on Hugging Face SpacesPaste a game and get Stockfish evaluations, move labels, accuracy and ACPL — plus coach notes whose selected numerical and move-label claims are checked against engine output before display.
Immortal Game regression · 18.Bd6
Source (opens in a new tab)74.3% → 50.0% · Blunder
Depth-16 Stockfish fixture; 288 centipawns lost; engine preferred Re1.
Single-file agentic framework
v0.4.0 · pre-1.0An agentic AI framework — tools, memory, RAG, structured output and multi-agent routing — written as one readable Python file with exactly two dependencies.
Offline test surface
Source (opens in a new tab)488 test functions
Counted from three public test modules; CI spans Python 3.9–3.13.
Supporting work
Multi-agent adaptive tutor
2026A self-hostable multi-agent language tutor that turns conversation into per-skill mastery evidence, schedules review with FSRS and plans the next most valuable thing to practise.
Evaluation design
Source (opens in a new tab)5-arm ablation · 30 ADRs
Deterministic fake-LLM mode; no study with real learners and CEFR is not human-validated.
Context-engineering experiments
2026An experimental platform for testing whether external cognitive memory can shrink prompt context in long conversations without losing task performance — with controlled baselines, ablations and a failure taxonomy.
Controlled comparison design
Source (opens in a new tab)5 conditions · 4 ablations
Current observations use one seed/length tier, an estimated token counter and no model in the loop.
Mathematics curriculum
2026104 Jupyter notebooks across 14 modules — from linear algebra and calculus to information theory, statistical learning theory and the mathematics behind transformers, LoRA and diffusion — each with a from-scratch NumPy or PyTorch implementation.
Verified curriculum footprint
Source (opens in a new tab)104 notebooks · 14 modules
Counted from numbered curriculum folders; excludes template and test-fixture notebooks.
Interactive knowledge base
202643 in-depth guides connecting classic data structures and algorithms to where they appear in AI/ML/LLM systems, with in-browser Python via Pyodide/WebAssembly and step-by-step algorithm animations.
Verified repository content
Source (opens in a new tab)43 guides · 43 practice pages
Repository count; the currently deployed GitHub Pages build still serves the earlier 31-guide version.
Proposed study design. No experimental findings or public thesis artifact are claimed on this site.
02Proposed research direction
Model architecture gets most of the attention, but balance, diversity, size, augmentation and preprocessing can set the ceiling on what training achieves. This proposed study would treat dataset structure as the independent variable and ask how each dimension changes what a model learns and how well it generalises.
Dataset characteristics
Balance, diversity, size, augmentation, preprocessing
Model architecture
Convolutional networks and Transformers
Training
Fixed budgets and controlled schedules
Evaluation
Held-out and shifted test sets
Generalization
The gap between fitting and understanding
Variables
03Repository writing
Architecture documents, security models and evaluation write-ups—not a fabricated publication history. Each entry opens its source repository.
LLM systems
Model & design
Eight memory types, provenance on every record, and a lifecycle in which nothing is silently overwritten: candidate → validated → active → superseded, expired, archived or quarantined.
LLM evaluation
Architecture note
The design behind a chess pipeline whose selected numerical and move-label claims are checked against Stockfish output, with zero-secret defaults and a runnable fallback.
Agent safety
Security model
A candid security model: tool authorization as a policy layer rather than a prompt, prompt fencing and its limits, and the known non-boundaries — in-process tools, thread timeouts, routing.
04About
I’m Nirav Vaghasiya, an AI/ML engineer and researcher. My public work focuses on the infrastructure around language models: memory, retrieval, tool use, orchestration and the checks that keep generated output grounded.
Across my repositories, I put deterministic components first and generative ones last. I document limitations beside the evaluation that supports each claim.
05Contact
Email is the fastest route. LinkedIn works too, and the code is always open for inspection.