Skip to work
Taimour Abdul KarimGet in touch
Lahore, Pakistan · 31.55°N 74.34°E

Taimour Abdul Karim · Data Scientist · AI Engineer

I build AIthat ships.

LLM evaluation, agentic systems, and production GenAI for teams from London to California. I measure models before I trust them: red-teaming, hallucination benchmarks, judge calibration, regression gates. Currently building Bryge.io at Datality while pursuing an MSc in AI at LUMS.

See the workRésumé
Taimour Abdul Karim, photographed in Lahore
Fig. 1 · T. A. KarimLahore, Pakistan
3+Years shipping AI3,000+GitHub stars earned20k → 900 linesBryge agent, rewritten+20%Accuracy lift on Llama 3.1

Shipped for Datality / Bryge.io · Qult Technologies · BornGreat · LUMS

Who I am

I'm a data scientist and AI engineer. Three years in, I built Bryge.io end to end: you point it at a Postgres database you already own, and a tool-calling agent on AWS Bedrock answers questions in plain English, returning the SQL and the chart that prove the answer. My specialty is the part most teams skip: evaluation. I red-team my own apps, audit the LLM judges everyone else trusts, measure where a model's confidence stops matching its correctness, and put regression gates in CI so prompt changes can't silently break production. MSc in AI from LUMS; 3,000+ GitHub stars.

Now
Building Bryge.io at Datality (London, remote) · MSc AI at LUMS
Focus
LLM evaluation & safety, agentic systems, RAG, production GenAI
Base
Lahore, Pakistan · working across UK / US time zones

What I represent

Production over demos

If it can't survive real traffic, it isn't done. Everything I ship is dockerized, monitored, and built to be handed over.

Grounded over plausible

LLMs earn trust by citing sources. My RAG systems constrain every answer to retrieved context. No confident hallucinations.

Measured over claimed

“+20% accuracy” means a benchmark, not a feeling. Improvements get numbers, baselines, and reproducible runs.

Research is easy. Production is the test, and these systems passed it.

What I’ve built

56 case studies · 2023 to 2026

01 · Flagship · 2026 · Capstone · All 8 milestones built and measured

Self-Improving Agent Platform

The Closed Retraining Loop

Every LLM agent in production degrades or overspends, and the fix is always a human: someone reads the traces, notices the pattern, rewrites the prompt, swaps the model. That human is the bottleneck and the reason most agent projects stall. Nobody ships the loop that does it without them, because closing it means solving the hard half first, an objective quality signal you can promote on.

Read the case studySource

System notes

  • Execution-based oracle: SQL answers are scored by running them, not by asking another model, which is what stops the loop reward-hacking its own judge
  • The CI gate flips 92% green to 67% red on a deliberately worse prompt, and it caught a citation collapse an execution-only gate would have missed
  • The on-device reranker fine-tune was rejected correctly: it lifted recall@3, recall@10 and nDCG but regressed rank 1 from 0.357 to 0.314, so the dominance gate declined it
  • Every guardrail fired on a live run, including an injection blocked at $0.00 before it reached the model

AWS Bedrock · FAISS + BM25 · FastAPI · Next.js · MLX / torch on-device · SQLite tracing

022026Clinical LLM Bias Audit · The Geographic Disparity IndexA clinical assistant that tells a Boston patient to come in for a visit but tells an identical Lagos patient to manage it at home is an equity failure with real stakes. Standard accuracy benchmarks cannot see geography-driven or name-driven disparity, so it goes unmeasured.AWS Bedrock · OpenAI · Groq · FastAPI · Streamlit · Wilcoxon / BCa bootstrap
032026SEC RAG Analyst · Hybrid Retrieval over 10-K Filings10-K filings are the hard case for RAG: 100-plus pages, dense tables, repetitive boilerplate, and cross-references, where naive embed-everything plus top-k cosine retrieves plausible but wrong passages. Answers without provenance or measurement cannot be trusted in financial analysis.BM25 · FAISS · BGE · Cross-encoder · Claude · FastAPI

Don't take my word for it. The code is public, and 2,400+ developers starred it.

Proof, in public

github.com/tkarim45 ↗

Most-starred repository · Jupyter Notebook

Beginner-Data-Science-Projects

A curated collection of hands-on data science projects. The most starred repo in my portfolio, ranked by oosmetrics Top 3 in Notebooks and Top 10 in Educational by growth.

0

stars · 564 forks

clinical-llm-bias-auditReproducible fairness-audit framework for clinical LLMs: the Geographic Disparity Index, with from-scratch statistics.new
sec-rag-analystProduction-style RAG over SEC 10-K filings: hybrid BM25 plus dense retrieval, RRF fusion, cross-encoder rerank, cited answers.new
credit-default-mlopsEnd-to-end MLOps: DVC, MLflow, a CI quality gate that blocks bad models, drift detection, and Prometheus-instrumented serving.new
everytongueTrain a neural translator for any language from a spreadsheet. Low-resource NLLB-200 recipe, pip-installable.new
site2botTurn any website into a fully offline local chatbot. No API keys, no cloud, about 600 lines of Python. Published on PyPI.new

487 followers · 74 public repositories · contributing since Jul 2021

Where I’ve shipped

4 roles · 2022–present

Jan 2024 – Present

London (Remote)

Datality · Bryge.io

Fullstack AI Engineer

  • Architected and built Bryge.io end to end: you point it at a Postgres database you already own, it works out the schema at runtime, and a tool-calling agent on AWS Bedrock answers questions in plain English, returning the SQL and the chart that prove the answer.
  • Rewrote the agent around making an unknown schema legible to the model rather than working around it, replacing roughly 20,000 lines of scaffolding with about 900. On a third-party schema it makes no SQL errors and joins the right tables on 9 of 10 questions with no table names in the question.
  • Made an MCP server the sole database gateway: the backend carries no database driver in its request path, so every read and write, including the agent's own SQL against customer databases, passes one audited chokepoint.
  • Took foreign keys from the Postgres system catalog after finding that the read-only login customers actually grant sees none of them through the standard route, which had been silently emptying the schema graph.
  • Instrumented the cost path to ~3.5¢ per question on a real third-party schema, routing Claude Haiku for chat against Sonnet for reports and pricing cache reads and writes separately.
  • FastAPI, Next.js, React, PostgreSQL, pgvector, AWS Bedrock (Claude), MCP, CloudFormation, S3 + CloudFront.

Aug 2023 – Dec 2024

Lahore

Qult Technologies

AI/ML Engineer

  • Fine-tuned Llama 3.1 8B for text classification: +20% accuracy over the base model on production NLU tasks.
  • Took models from notebook to production with zero manual steps: end-to-end pipelines on AWS SageMaker and Lambda.
  • APIs, NLP, LLMs, Docker, AWS, TensorFlow, GitHub.

Sep 2022 – Feb 2023

California (Remote)

BornGreat

Data Scientist

  • Built NLP + LLM sentiment analysis over social media streams, turning raw posts into supply, demand, and competitor signals the team acted on.
  • Shipped a FastAPI service for real-time model inference, wired to GitHub Actions so every merge deploys itself.
  • Selenium, APIs, NLP, LLMs, FastAPI, Docker, TensorFlow.

Jun 2021 – Aug 2021

Lahore

Spyresync

Python Developer (Intern)

  • Built interactive Django dashboards over datasets I cleaned and feature-engineered myself.
  • Integrated scikit-learn models into web apps to ship predictive features to real users.
  • Python, APIs, GitHub, Django, Scikit-learn.

Stack & credentials

Working stack

Data Science / ML
Python · SQL · Pandas · Scikit-Learn · PyTorch · TensorFlow · SciPy · Causal inference · Statistical analysis
GenAI / LLM Engineering
AWS Bedrock (Claude) · OpenAI APIs · LangChain · LangGraph · RAG architectures · Agent tool design · Fine-tuning (Llama, NLLB) · pgvector / FAISS / ChromaDB
LLM Evaluation / Safety
DeepEval · RAGAS · LLM-as-judge calibration · Red-teaming (OWASP LLM Top 10) · Hallucination benchmarks · Fairness auditing · CI regression gates
Infra / MLOps
FastAPI · Docker · AWS · DVC · MLflow · Prometheus · Grafana · PostgreSQL · pgvector · GitHub Actions · MCP

Education

  • MSc. in Artificial Intelligence

    Lahore University of Management Sciences (LUMS)

    Sep 2024 – Jun 2026

  • BSc. in Data Science

    National University of Computing and Emerging Sciences

    Sep 2020 – Jun 2024

Recognition

  • 2nd Position, Genesis Hackathon Dubai
  • Exceptional Deputy Head, SOFTEC
  • Data Manipulation, DataCamp
  • Joining Data with Pandas, DataCamp

Open to AI/ML engineering roles, remote or hybrid

Let’s build something that ships.

taimour.a.karim@gmail.com

or +92 326 1127700 · usually replies within a day

© 2026 Taimour Abdul KarimDesigned & built in Lahore