Shipping applied AI · Remote / Islamabad

Building thesystems layer betweenLLMs and reality.

I'm Yasir Raza — an AI Engineer building production-grade agent systems, multimodal voice interfaces, fine-tuned models, and MCP tooling. From ambiguous prototype to measurable ROI.

Get in Touch Selected Work

10.7K+

PYPI DOWNLOADS

Across notebookllm and notebookllm-mcp packages

7,275

TESTS WRITTEN

Across Kausal and the Archon 2.0 platform

4.5×

TTS SPEEDUP

bfloat16 + CUDA Graphs — RTF 0.59 → 0.13

13th

AGENTBEATS RANKING

/100+ overall · 49.7% avg · 85.1% defense

PythonPyTorchLangGraphUnslothFastAPIDockerLLamaIndexLoRA/PEFTMCPRunPodPythonPyTorchLangGraphUnslothFastAPIDockerLLamaIndexLoRA/PEFTMCPRunPod

01 / About

A builder at the intersection of research and shipping.

I've spent the last 3+ years turning applied ML research into systems that process real data — thousands of AI tools verified for a recommendation engine, job-posting intelligence at scale, a 7,275-test production agent platform, and PyPI packages with thousands of downloads.

My work spans the full stack of applied AI: designing agent architectures with LangGraph and pure-Python frameworks, building RAG pipelines, deploying voice agents on serverless GPU infrastructure, and shipping security platforms measured by 2,000+ tests. I'm obsessed with the parts that separate demos from products — latency budgets, evaluation harnesses, safety filters, and reproducible deployment runbooks.

Previously building data pipelines and AI systems at Teamlift (US remote). Graduated B.Sc. Software Engineering (NUML, Aug 2026). Currently open to AI engineering roles, freelance projects, and interesting collaborations in the agentic AI space.

CORE FOCUS

LLM Agents & Voice AI

SECONDARY

Fine-tuning, RAG, MLOps

BASED IN

Islamabad, Pakistan

AVAILABILITY

Open for freelance

02 / Capabilities

What I build.

Specialized across the full AI engineering lifecycle — from whiteboard to deployment to monitoring.

Agentic AI Systems

Custom autonomous agents using LangGraph & LangChain. Multimodal inputs, tool-use, memory, and complex task decomposition for real-world automation.

LLM Fine-Tuning

Domain-specific model adaptation using PEFT, LoRA, QLoRA via Unsloth. Specialize models on proprietary data without breaking the bank on compute.

RAG & Data Pipelines

Production-grade Retrieval Augmented Generation. Semantic chunking, hybrid retrieval, and vector search to ground LLMs in enterprise knowledge.

Voice & Multimodal Agents

Real-time voice/video assistants with cloned voice, emotional TTS, and multimodal understanding. Built for accessibility and personalized UX.

MCP Server Development

Model Context Protocol servers enabling LLMs to interact with tools and environments. NotebookLLM has 10.7K+ PyPI downloads.

MLOps & Deployment

Dockerized, CI/CD-driven deployment. Serverless inference, FastAPI wrappers, and robust monitoring for production AI systems.

03 / Selected Work

Things I've shipped.

Open-source packages, agent systems, and production deployments. Each one taught me something I didn't know.

001 / FLAGSHIP

Kausal

Production AI agent platform (private): extended the ii-agent framework in pure Python (no LangGraph) with custom RAG, layered memory, full OpenTelemetry observability, realtime voice/video via LiveKit, LiteLLM multi-model routing, agent skills, and S3 storage.

→ 7,275-test safety net across 16 test categories; architecture-level observability with self-hosted SigNoz.

7,275 testsOTel tracing + metricsLiveKit voice/videoLiteLLM routing

PythonFastAPIOpenTelemetryLiveKitLiteLLMS3PostgreSQL

001 / FLAGSHIP

NotebookLLM

Open-source Python package + MCP server that makes Jupyter notebooks efficient for LLM workflows — agent-optimized text output, token budgeting, streaming for 10MB+ notebooks, and cell execution across 8+ notebook formats.

→ 10.7K+ PyPI downloads. Used by Claude Desktop, Cursor, and VS Code workflows via MCP.

10.7K+ PyPI downloads8+ notebook formatsUp to 80% token reductionMCP server built in

PythonPyPIMCPClicktiktokenasyncio

001 / FLAGSHIP

Math2Code

A math LLM translating LaTeX into executable, verified SymPy code — trained with rule-based RLVR (GRPO on TRL), judged by a competition-faithful execution harness with a SHA-256 frozen split and sandboxed end-to-end.

→ Rebuilt a data-leaked capstone into a measurement-first system on a <$100 budget.

397/397 gold solutions verifiedSandbox pool: 389 exec/s22,796 deduplicated samples<$100 total budget

PyTorchTRL / GRPOSymPyDockerHuggingFace

001 / FLAGSHIP

Cerberus

Stealth browser infrastructure for AI agents — a self-hostable control plane pairing a source-level stealth Chromium with token-efficient orchestration, a multi-agent isolation gateway, and persistent session memory.

→ Twelve build phases verified end-to-end with committed evidence files.

60-tool MCP bridge86 automated verification checksMulti-agent isolation gateway

PythonMCPChromiumDockerNode.js

001 / FLAGSHIP

Forge

The API compiler for the agentic era — one OpenAPI 3.x spec in, production-grade SDKs in 10 languages, CLIs, MCP servers, Agent Skills, Terraform providers, and docs out. Open-source (MIT) and air-gapped.

→ TypeScript SDKs compile strict on Stripe (~3K files) and Cloudflare (~14K files) real specs; 13 compile gates in CI.

17 output types10 SDK languagesByte-stable outputBreaking-change gate

TypeScriptPythonRustGoOpenAPI

002

Archon 2.0

The agent-security platform evolved from Archon's 13th-place (49.7%) AgentBeats run — the only open platform where an adaptive attacker and a measurable defense fight in the same loop: 8-layer defense pipeline, 222-probe corpus across 10 attack packs, runtime defense proxy (OpenAI-compatible), MCP security scanning, and full OWASP Agentic Top-10 (ASI01–ASI10) coverage.

Competition-proven lineage (13th/100+, 85.1% defense) now a production platform: 2,295 tests, 93% coverage gate, AgentDojo v1 benchmark published, Homebrew + npm distribution.

From 13th/100+ AgentBeats agent2,295 tests222 probes · 10 packsOWASP ASI01–10 full

PythonA2A ProtocolGoogle ADKOpenTelemetryPostgres

003

Chatterbox-Optimized

The most optimized fork of Resemble AI's Chatterbox TTS — bucketed CUDA Graphs for autoregressive inference, a zero-hallucination pipeline selecting lowest-WER audio via in-memory Faster-Whisper, and long-text chunking.

Production serverless engine deployed to RunPod with reproducible benchmarks.

4.5× faster · RTF 0.59 → 0.13Whisper WER validationServerless-ready

PyTorchCUDA GraphsFaster-WhisperRunPodDocker

004

Archon

Adversarial AI agent security framework built for the Lambda × Berkeley RDI AgentBeats Security Arena — GOAT-style adaptive attackers battle 7-layer defense gateways across prompt injection, exfiltration, and jailbreak scenarios.

13th / 100+ overall (49.7% avg win rate); the scenario-agnostic attacker/defender architecture generalized to unseen private-leaderboard scenarios.

13th / 100+ overall · 49.7%6th Defense (Blue Team) · 85.1%18th Attack · 14.4%

PythonA2A ProtocolGoogle ADKRed-teaming

005

PRISM-Bench

Cultural intelligence benchmark for AI systems exploring pluralistic reasoning and identity-specific modeling.

Adds evaluation and benchmark literacy to a portfolio heavy on shipping systems.

Cultural Intelligence (CQ) benchmark

PythonBenchmarkingEvaluationResearch

04 / Experience

The path here.

Junior Data Scientist

Jan 2024 — Dec 2024

Teamlift · US Remote

Scraped thousands of AI tools from public tool directories; built an automated verification pipeline (HTTP status checks, screenshot capture, freshness validation) powering Teamo, Teamlift's recommendation engine; preprocessed and normalized thousands of tool records into model-ready features.

Data Science Intern

Jul 2023 — Dec 2023

Teamlift · US Remote

Built an automated pipeline collecting thousands of job postings from dice.com with deduplication; extracted skill-demand signals and trained time-series models to forecast skill demand from job descriptions.

05 / Certifications