33.6844° N 73.0479° E ISB --:--

AI / ML ENGINEER APPLIED AI SYSTEMS FULL-STACK BUILDER

HASNAAT HUSSAIN

CURRENT THESIS

Building intelligence that survives contact with the real world.

00 / THESIS NOISE IN → SIGNAL OUT

Models in motion. Reliable systems. Useful products.

I turn research-grade ideas into dependable AI products: model behavior, inference constraints, backend systems, and the interface a person actually uses.

The advantage is not another tool list. It is seeing the entire product as one connected instrument.

01 / MODELLearn the signal

02 / INFERMake it reliable

03 / SHIPProduct surface

01 / SELECTED TRANSMISSIONS SELECTED SYSTEMS / LIVE + OSS

Work that moved from idea to interface.

Selected products where machine intelligence is part of the operating system, not an ornamental feature.

RETENTION SIGNAL / LIVE MODEL CONFIDENCE 92.4%

01 / 06

AI-POWERED B2B SAAS

Anchoryn

Find the churn signal before the customer leaves.

A retention operating surface combining behavior, account health, support pressure, billing signals, and AI-guided interventions for founders and customer-success teams.

ROLE: Product / ML / Full stack
SYSTEM: PyTorch, FastAPI, PostgreSQL, Next.js
FOCUS: Risk prediction & actionability

Open live product↗

RETROFORGE.EXE_ □ X

WEB TIME MACHINE

ERA SELECTORX

WINDOWS 98
GEOCITIES
CRT TERMINAL
MAC CLASSIC

PROCESS COMPLETE 8 ARTIFACTS GENERATED █

AUTHENTICALLY INCORRECT / 1998–2026

02 / 06

GENERATIVE WEB TOOL / MONOREPO

RetroForge

Send any website backwards through the internet.

A nostalgia-heavy generator that converts public sites into era-correct artifacts: XP, Windows 98, GeoCities, CRT, Netscape, ANSI/BBS, and early Blogspot.

ROLE: Concept / System / Frontend
SYSTEM: Next.js, TypeScript, Python, Tailwind
FOCUS: Generative art direction

Open live product↗

01 / INPUT NICHE Objects for slow mornings

02 / THEME

Editorial commerce

03 / COPY SYSTEM

04 / DEPLOY STORE LIVE 00:57

03 / 06

LLM TOOLING / COMMERCE

StoreCraft AI

From a niche to a configured Shopify store in under a minute.

An AI store builder generating themes, product copy, email flows, ad creative, and policies before deploying the finished storefront.

ROLE: Product / AI workflows / Full stack
SYSTEM: Next.js, Node.js, OpenAI, Shopify API
FOCUS: Generative operations

Open live product↗

AGENT RUN / 0142 GATED / 0.84

PRODUCTIONOS / EVIDENCE LOOP RUNNING

CLAIM → EVIDENCE → GATE PRODUCTIONOS / OSS

04 / 06

AI AGENT INFRA / TYPESCRIPT

ProductionOS

Give coding agents a memory for what actually shipped.

An evidence-driven operating layer for AI coding agents: capture claims, inspect the diff, run the right checks, and leave a durable trail before a change reaches production.

ROLE: Architecture / Product / Full stack
SYSTEM: TypeScript, Next.js, SQLite, GitHub Actions
FOCUS: Agent reliability & developer velocity

Open GitHub repository↗

CPU INFERENCE / FP16 LATENCY 7.8 MS

PREPROCESS → EXECUTE → DECODE NN_INFERENCE / C++17

05 / 06

INFERENCE RUNTIME / C++17

nn_inference

Make a trained model feel native on the machine it runs on.

A CPU-first ONNX runtime experiment focused on predictable latency, small memory surfaces, and the unglamorous details between a model artifact and a useful prediction.

ROLE: Runtime design / Optimization
SYSTEM: C++17, ONNX Runtime, SIMD, CMake
FOCUS: Inference performance & portability

Open GitHub repository↗

LOCAL MULTIMODAL RAG GROUNDING / 96%

YOU

What changed in the brief?

OMNI / LOCAL

The scope moved from a feature list to an evidence trail.

[01] BRIEF.PDF [02] NOTES.MD

ASK → RETRIEVE → GROUND OMNICHAT / CASE STUDY

06 / 06

MULTIMODAL RAG / PRODUCT SYSTEM

OmniChat

A quiet, local-first interface for asking better questions of messy context.

A sanitized case study for a multimodal retrieval system that keeps documents, screenshots, and citations close to the user instead of hiding the hard parts behind a chat bubble.

ROLE: Applied AI / UX / Full stack
SYSTEM: Python, embeddings, multimodal retrieval, React
FOCUS: Grounded answers & trust

Open GitHub case study↗

02 / OPEN-SOURCE PATCH BAY FAULT → INTERVENTION → MERGE

Fixing the layer underneath.

Production-grade AI and data infrastructure improves one precise upstream intervention at a time.

PYDANTIC AIMERGED

6098

Amazon Bedrock tool-result co-location across model families.

Prevented validation failures during agent conversations by handling provider-specific content constraints correctly.

BUG FIX / COMPATIBILITY↗

SQLGLOTMERGED

7807

Correctly parsed parenthesized multi-expression queries.

Resolved a parser edge case where expression tuples were interpreted as subqueries.

PARSER / CORRECTNESS↗

PLOTLY.JSMERGED

7768

Small-number tick labels in exponential notation.

Fixed axis formatting across chart types for values near the exponential threshold.

AXES / FORMATTING↗

ULTRALYTICSMERGED

25153

RT-DETR detection limits across export and validation.

Honored max_det consistently in decoder exports and evaluation paths.

MODEL EXPORT / DETECTION↗

DATAFUSIONMERGED

24484

CASE evaluation with custom physical expressions.

Preserved custom column expressions through physical projection planning.

QUERY ENGINE / CORRECTNESS↗

MLFLOWMERGED

25807

Keep Strands autologging stable across foreign spans.

Guarded trace-id handling so MLflow integrations stay compatible with OpenTelemetry spans they did not create.

OBSERVABILITY / AGENTS↗

Open review queue. 08 ACTIVE UPSTREAM PROPOSALS

LITELLM / 32452Support drop_params for embedding dimensions across Azure and OpenAI-compatible calls.OPEN ↗ LITELLM / 31070Honor drop_params in the Anthropic pass-through endpoint.OPEN ↗ PANDAS / 66209Fix resampling with CustomBusinessDay and a right-closed window.OPEN ↗ PHOENIX / 15584Improve tracing behavior for the next generation of LLM workflows.OPEN ↗ GRAPHITI / 1604Keep temporal graph retrieval consistent as context grows.OPEN ↗ VLLM / 49639Refine serving behavior for production inference workloads.OPEN ↗ BENTOML / 5671Make model-serving validation safer for typed return values.OPEN ↗ CREWAI / 6604Strengthen agent orchestration behavior around tool execution.OPEN ↗

03 / CAPABILITY SPECTRUM MODEL WEIGHTS → USER SURFACE

One connected instrument.

Depth in model behavior, fluency across the production stack, and a bias toward systems that stay legible after launch.

BAND / 01

Model systems & inference

Python, C++, PyTorch, Transformers, ONNX Runtime, quantization, PEFT, evaluation, SIMD

BAND / 02

LLM & applied AI

RAG, embeddings, multimodal retrieval, LlamaIndex, LangChain, Ollama, OpenCV, YOLO, XGBoost

BAND / 03

Backend & data

FastAPI, Node.js, PostgreSQL, Redis, Celery, Kafka, REST, GraphQL, background jobs

BAND / 04

Frontend & product

React, Next.js, TypeScript, Tailwind CSS, auth, interaction design, accessibility, Vercel

BAND / 05

Delivery & open source

Docker, Linux, GitHub Actions, CMake, Ninja, pybind11, Protobuf, OpenTelemetry, pytest, Vitest

04 / FIELD NOTES ISLAMABAD / PAKISTAN

SUBJECT HH / FRAME 02

BACKGROUND / DIFFERENTIATOR

A builder obsessed with useful intelligence.

I am an applied AI engineer and full-stack builder based in Islamabad. I work from model behavior to the product surface: retrieval, inference, backend systems, and interfaces that make complex tools feel calm.

I also work upstream, finding precise failures in production-grade AI infrastructure and fixing them at the source.

JUL 2026 — PRESENT

NATIONAL CENTER FOR PHYSICS / COE AITEC

ML / DL Intern

Deep-learning research in Islamabad, applying modern machine-learning architectures to real research problems.

2023 — 2027

QUAID-I-AZAM UNIVERSITY

BS Electronics / AI focus

Building depth in machine learning, computer vision, inference systems, and the mathematics underneath reliable software.

05 / LIVE DEVELOPMENT SIGNAL SNAPSHOT / 2026-09-14

Work in the open.

A live readout of public repositories and contribution cadence. Static evidence remains when the network does not.

37PUBLIC REPOSITORIES

11GITHUB FOLLOWERS

42TRACKED PRS

06MERGED UPSTREAM

—PROFILE VIEWS

00:00TIME ON PAGE

CONTRIBUTION DENSITY / 52 WEEKSLIVE SIGNAL

Full profile ↗

CONTACTING GITHUB SIGNAL...

QUIETACTIVE

06 / ESTABLISH A LINK AVAILABLE FOR SELECT WORK

HAVE A HARD PROBLEM?

Bring the signal.

hasnaat.hussain.2@gmail.com