Background
I build production LLM systems - semantic search and retrieval pipelines, held to what the data behind them can actually support.
I'm an AI/ML engineer with an MSc (Merit) in Artificial Intelligence from the University of Bradford and a first-author Springer paper. My work sits where applied research meets production: measuring a system honestly, then making it fast enough to put in front of real users.
Right now I build and run Jobzyl on my own: a live job search aggregator over 29 integrated sources and a 3.4M row Postgres index, with semantic CV-to-posting matching in pgvector. The half I care about is what it refuses to say. Pay renders in the period the employer quoted or not at all, and a posting that names no recognised skill reports as too thin to score rather than as a confident 0%.
Before that, at Outlyst, I cut mean call latency 54%, from 2.4s to 1.1s across 2,100+ calls, by profiling async I/O and restructuring connection pooling. For my dissertation I built FinLaw-UK, a graph-augmented RAG system for UK financial regulation where a Neo4j knowledge graph resolves every citation before the answer ships, and refuses rather than answer when it cannot. After submitting, I re-measured my own evaluation and found two of the reported figures were regex shape-checks rather than correctness measures; I published the correction and a report on which numbers reproduce.
I'm interested in roles where I can keep pushing that boundary - semantic search and retrieval at scale, LLM infrastructure, or high-throughput ML systems.
Spoken languages: EN fluent · UR native · DE A1
Education
MSc in Artificial Intelligence (Merit)
University of Bradford
Sep 2024 - Sep 2025 · Bradford, UK
Dissertation - FinLaw-UK: A Graph-Augmented Retrieval Chatbot for Reliable and Transparent UK Financial Regulation
Focus - Modules - Artificial Intelligence and Data Science (79), Business Data Analytics (79), Responsible AI: Ethics, Law and Governance (75)
BS in Bioinformatics
COMSATS University Islamabad
Sep 2020 - Jul 2024 · Islamabad, Pakistan
Dissertation - AI-Assisted Analysis and Prediction of At-Risk Diabetic Individuals
Focus - Predictive analytics, interpretability, biological impact modelling
Technical skills
Pick one to see the work behind it.
Machine learning, NLP & LLMs
Engineering & data
Infrastructure & quality
Where I've built
Engineering roles and research posts, most recent first - what I built and shipped at each.
Founder & Sole Engineer
Jobzyl
- Build and operate a live, public job search aggregator end to end: 29 provider integrations and 6 applicant tracking systems, a 3.4M row Postgres index across 26 countries, FastAPI backend, Next.js frontend and AWS deploy pipeline.
- Shipped semantic CV-to-posting matching in production: 384-dimension multilingual sentence embeddings over pgvector against the full corpus, gated so a posting naming no recognised skill reports as too thin to score rather than a confident 0%.
- Built the LLM layer on Anthropic Claude (CV scoring, cover letter, interview prep), with prompt-injection defences on every call, per-user quotas and documented fail-open behaviour.
- Trained a pay regression model against the shipped salary benchmark as baseline, split by employer group to prevent leakage: MAE 27,420 to 24,001, MdAPE 18.9% to 16.4%. Reported per country and seniority, and held back from deployment.
- Rebuilt search ranking as a weighted Postgres full-text function with a title-relevance layer, after measuring that 27.4% of returned results had none of the user's query terms in the job title.
- Hardened the platform: Fernet field encryption of CV text, row-level security across 23 tables, GDPR export and deletion, behind 1,950 automated tests and 11 CI build gates.
AI / Machine Learning Engineer
Outlyst
- Engineered and optimized the inference architecture for an autonomous AI voice agent (Retell AI, FastAPI), handling 2,100+ outbound calls.
- Enhanced agent capabilities to detect gatekeepers and schedule callbacks rather than dead-ending the transfer.
- Conducted rigorous backend profiling to isolate inefficient asynchronous I/O and connection pooling, driving a 54% reduction in systemic latency (2.4s → 1.1s).
- Built an internal micro-CRM with automated contact-extraction pipelines, removing external CRM licensing costs.
- Built VoiceFlow, a FastAPI service running Whisper large-v3 in-process over the 2,100+ call recordings, threading each export so blocking downloads and torch inference stay off the asyncio event loop.
Research Assistant, Graph-Augmented LLM Engineering
University of Bradford
- Engineered FinLaw-UK, a Retrieval-Augmented Generation architecture integrating Mistral 7B-Instruct with a Neo4j knowledge graph for domain-aware regulatory retrieval. This work was also my MSc dissertation project - the Education and Experience entries describe the same system, not two.
- Built a hybrid retrieval pipeline: BM25 sparse retrieval fused with BGE-small dense embeddings via reciprocal rank fusion, then cross-encoder re-ranking, with graph-grounded citation verification over Neo4j.
- Evaluated on a 110-item regulatory benchmark I built and released: 80 factual questions, 20 document tasks, 10 case scenarios. Reproducible figures are 0.76 faithfulness and 0.74 answer relevance (RAGAS) with legal completeness at 0.68.
- Re-measured the submitted evaluation and found the reported source-accuracy and citation-quality figures were regex shape-checks rather than correctness measures: the scorer returned a flat 0.85 for 103 of 110 rows, and the true graph-verified citation rate was 3 in 110. Published the correction and a measurement-integrity report stating which figures reproduce.
- Built reproducible experimental pipelines with structured evaluation protocols, extending RAGAS with a custom legal-completeness metric that holds at 0.68 across both evaluation tracks.
Research Intern, Data Science
COMSATS University Islamabad
- Benchmarked 11 ML classifiers for diabetes risk on 253,680 CDC BRFSS records, resampling confined to the training folds; Random Forest led at 93% accuracy and performed best on ROC-AUC and sensitivity.
- Deployed predictive models via REST APIs with correlation-driven risk-factor analysis surfacing the drivers behind each score.
- Built a production web interface (React.js + Flask) for real-time clinical risk scoring.
- First and corresponding author on a Springer conference paper; presented at ICSMAI 2024, Saidia, Morocco.
Projects & case studies
Selected projects spanning AI research, ML systems, and full-stack engineering - each backed by a full case study on the architecture and the decisions behind it.
- Full-stack
Jobzyl
One search across 29 job sources, with semantic CV matching
3.4M rows in the index · 29 sources integrated
- Research
FinLaw-UK
Graph-augmented RAG for UK financial regulation
0.76 ragas faithfulness · 0.74 answer relevance
- Engineering
Autonomous Voice Agent
Cutting voice-agent latency 54%, from 2.4s to 1.1s
54% latency reduction · 2,100+ calls handled
- ML
DiabetesSense
93% accurate diabetes risk screening from an 11-model benchmark
93% accuracy (random forest) · 11 models benchmarked
- Engineering
VoiceFlow
Retell call exporter with local Whisper transcription
local whisper large-v3 · SSE live job progress
- Research
Sleep Efficiency Predictor
Published ML research served as a lifestyle-factor prediction app
0.8569 r² · random forest · 452 records · 15 features
Notes & deep dives
What I learned building the systems above, written up while it was still fresh - evaluation methodology, latency profiling, and the decisions that did not survive contact with real data.
- Deep dive
49 of 50 salaries were not salaries: what a job aggregator is allowed to claim
Fetching a million job listings is an afternoon. Deciding what you are willing to say about them - the pay, the dedup, whether the role still exists - is the whole engineering problem.
Aug 2026 · 7 min read
- Deep dive
Three scores, one CV: keyword, semantic, and the one that leaves
Keyword extraction in the browser, cosine similarity in pgvector, and Claude for the writing. Three paths rather than one, because they differ in what they can see and how far your CV has to travel.
Aug 2026 · 5 min read
- Deep dive
Gold answers and required citations: designing a benchmark for legal RAG
The FinLaw-UK pipeline was scored against a benchmark I had to build first. Writing 110 legal QA items taught me more about RAG failure modes than building the pipeline did.
Aug 2026 · 5 min read
- Deep dive
Ranking is the product: 27.4% of results had nothing to do with the query
Retrieval stopped being the hard part once the index got big. Ordering took over, and the first version shipped a quarter of its results with none of the query in the job title.
Jul 2026 · 5 min read
Publications
Peer-reviewed work presented at international venues.
Graph-augmented retrieval, LLM faithfulness evaluation, systems optimisation for high-throughput ML pipelines, and interpretable predictive modelling for clinical applications. Open to funded PhD positions in the EU and UK.
ORCID 0009-0000-7873-4977Get in touch
I'm available now for full-time AI/ML engineering and research roles, and open to funded PhD positions in the EU and UK. If you're hiring for an AI/ML team, supervising research, or just want to compare notes on retrieval systems, email is the fastest way to reach me.
hammadahmad.ml@gmail.comAvailable for workBradford, UK · open to relocation · replies within a day