Muhammad Tauha

AI Engineer specializing in Agentic LLM Systems, Structured Outputs, and Local Models.

Based in: Lahore, PakistanMost recently: Data Engineer (LLM Systems), Victreat Health Tech

Muhammad Tauha is an AI Engineer who builds production-grade LLM systems on private infrastructure, notably increasing document classification recall from 36% to 95% using quantized open models. He maintains high-traffic public APIs and open-source MCP servers, focusing on deterministic structured outputs and efficient RAG architectures.

Experience

Data Engineer (LLM Systems)

Victreat Health TechIslamabad, PakistanworkApr 2026 – Sep 2026

Summary

Designed an LLM document classification pipeline for regulated, on-premises clinical data: the model classifies and deterministic code extracts, after testing showed LLM extraction invented values. Recall rose from 36% to 95%, measured on a golden set of 4K+ labeled documents against the old classifier. Benchmarked Qwen3, gpt-oss-20b and Gemma 4 on in-house faithfulness tests and shipped 4-bit quantized Gemma 4 on vLLM, fitting several instances in a 20GB GPU budget; checked the quantization quality cost on the same benchmark. Enforced correctness in two layers: vLLM JSON-constrained decoding for valid output and Pydantic schemas for field-level validation, with failures flagged for retry and root-cause review. Built an OpenAI-compatible load balancer (least-connections routing, 30s health checks) so services scaled across model instances by changing one environment variable. Built a document router on bge-m3 multi-vector embeddings (ColBERT-style late interaction over headings and lead lines), so short type-defining signals in long documents were not averaged away. Owned the ingestion layer feeding the RAG corpus (~480K documents a day, Pydantic-validated, Airflow into Elasticsearch); evaluated Skyvern, browser-use and Firecrawl for collection and added a human-in-the-loop handoff that resumes stalled jobs with session state intact. Built an in-house secrets-management service on PostgreSQL with RBAC and time-limited sessions.

What I did

Ran 4-bit quantized Gemma 4 on vLLM for regulated, on-premises clinical data.
Optimized throughput by using 4-bit quantization to fit multiple Gemma 4 instances within a 20GB GPU budget.

LLMQwen3gpt-oss-20bGemma 4vLLM4-bit quantizationJSON-constrained decodingPydanticOpenAI-compatible load balancerbge-m3multi-vector embeddingsColBERTRAGAirflowElasticsearchSkyvernbrowser-useFirecrawlPostgreSQLRBAC

Skills

Technical

FastAPI

Python

REST

RAG

vLLM

Pydantic validation

structured decoding

pytest

MCP

async I/O

PostgreSQL

DuckDB

OAuth 2.0

4-bit quantization

model benchmarking

LangGraph

embeddings

ChromaDB

Elasticsearch

incremental indexing

WebSocket

Redis

Docker

GCP Cloud Run

GitHub Actions

Airflow

OpenRouter

TypeScript

OpenTelemetry

Distributed Systems

System Architecture

Software Project Management

SDLC

Speculative Decoding

Load Balancing

SQL

Projects

psxdata-mcp

Lead Designer/Developeropen_source

Designed an 11-tool agent harness around a load step plus a single SQL tool, so computation runs in DuckDB instead of the model, keeping tool outputs small and calls few. Sandboxed tool execution (no file or network access, no extensions, 1GB memory, config locked before queries run), with a 200-row output cap and a 30s timeout via interrupt. Shipped locally (Docker image, Claude Code plugin, one-click desktop bundle) and hosted on Cloud Run behind a TypeScript Cloudflare Worker OAuth gateway with hourly access tokens and sliding refresh; 66 Python and 47 vitest tests, including an end-to-end OAuth flow.

MCPPythonTypeScriptOAuth 2.0DuckDBDockerClaude CodeCloud RunCloudflare Workervitest

View the code

Troubleshot.ai

CLI, Indexing, Retrieval, Generation and Evaluation Leadacademic

Agent that diagnoses and fixes developer-environment failures from system snapshots covering about 15 areas of the system; owned the CLI, indexing, retrieval, generation and evaluation while a teammate built orchestration. Built a Merkle-tree differential indexer (SHA-256 per leaf, skips unchanged subtrees, upserts or deletes by path) that cut ChromaDB re-indexing from 5-7 minutes to about 30 seconds on a 4-core CPU. Evaluated retrieval and generation separately; generation was the bottleneck, so moved to gpt-oss-120B for the largest faithfulness gain at about $10 total cost. Added human approval for commands, an abort interrupt that kills running tool calls, and regex redaction of secrets before snapshots reach the model.

LangGraphRAGWebSocketMerkle-treeSHA-256ChromaDBgpt-oss-120B

psxdata-api

Maintainerpersonal

Run a public API with 12 endpoints: 390K+ requests and 3.5K+ unique IPs a week, 0% 5xx errors. Traced a latency incident with OpenTelemetry and added a two-tier cache (in-memory LRU over Redis, single-flight locks, stale fallback), cutting p99 from 21.5s to 302ms and ending an upstream rate-limit incident.

FastAPIRedisOpenTelemetryLRU Cache

View the code

psxdata

Maintaineropen_source

Maintain a PyPI package with 31K+ downloads and 330+ tests that stays reliable against an unstable upstream through retries, dynamic parsing and a disk cache.

PythonPyPIDuckDBBigQuery

View the code

Final Year Project

personal

Focused on RAG and retrieval systems. Owned the retrieval component and optimized the re-indexing process.

Speculative Decoding Evaluation

personal

Setting up vLLM with a small draft model from the same family to evaluate acceptance rate and tokens per second on a public document dataset.

vLLMEAGLE-style draft headsn-gram prompt lookup

Education

B.E. Software Engineering

National University of Sciences and Technology (NUST)Software Engineering2022 – 2026

Thesis. Troubleshot.ai

Recognition

Certifications

Certified in Cybersecurity (CC)

ISC2

Google Cloud skill badges (BigQuery ML, Cloud Security Fundamentals, IAM, Text Prompt Engineering)

Google Cloud

Hire MuhammadReach out about a role, a contract or a conversation.For recruiters

Muhammad's twin is AI, it can make mistakes.

More people on Kanvis

Browse more people Create your own Kanvis