I build production agentic AI systems, then measure how they fail.
I treat LLMs as stochastic components that need deterministic guardrails, typed state management and a real evaluation harness — not as magic boxes. Most of my work is the second half of that sentence: building the system is the easy part, proving it behaves is not.
🟢 Status: BE Software Engineering, KIET, completed August 2026. Open to remote AI engineering roles. No visa sponsorship required. Also preparing Masters applications for Winter 2027/28.
| 🔬 Evaluation and reliability | 🏗️ Production systems |
|---|---|
| A 46-file behavioural test suite of 592 automated tests, 19 behavioural probes using paired before-and-after measurement, 60 per-turn traces across five personas, and three named reproducible LLM failure modes | Multi-agent systems with typed state, concurrent-write reducers, deterministic quality gates and streaming, recorded across 334 documented build and audit sessions |
| Evaluation Design · Behavioural Testing · Red Teaming · Token Sampling · LangSmith | LangGraph · Google ADK · FastAPI · Flutter · SSE Streaming |
| Project | Description | Stack |
|---|---|---|
| Rahzaan · AI Career Guidance (Source private · Case Study · Live Demo) | Deployed career counselling platform for Pakistani secondary school students. Maintained through an evaluation harness of 19 behavioural probes and 592 automated tests over a 24-node LangGraph pipeline, first response cut from 20–25s to 3.7s. 265-item psychometric assessment stack, 9-criterion reflection gate, live SSE streaming, 29-route Flutter frontend. The case study documents the failures as well as the build. | LangGraph FastAPI Flutter PostgreSQL |
| Autonomous Research Concierge | Multi-agent research pipeline. An intent synthesiser compacts chat history into a research directive before three search agents fan out concurrently. Per-call session isolation, and a deterministic code-based filter ahead of all tool access so the guardrail is structurally bypass-resistant rather than prompt-enforced. Deployed on Railway. | Google ADK Gemini API Docker Python |
| Applied LLM Workflow Research | Behavioural study of Gemini 2.5 Pro. Names three reproducible failure modes — Optimal Answer Trap, Final Synthesis Bias, Conversational Gravity — each with documented prompt sequences. Includes controlled sampling experiments, 1M-token context stress tests, and context-ordering trials showing a ~30% retrieval error rate when context follows rather than precedes the instruction. | Gemini API Streamlit Python |
| Crypto Accumulation Scanner | Scans 1,200+ cryptocurrencies daily across CoinMarketCap and MEXC Futures. Classifies price action into 7 accumulation signatures and applies a log-log regression volatility baseline derived from 2020–2024 market data, normalised by market-cap tier so outliers surface instead of size artefacts. | Google Apps Script JavaScript CoinMarketCap API |
| WordPy Solver | Autonomous Wordle solver reading game state by sampling tile pixel colour instead of OCR, since the corner region is pure background regardless of font or antialiasing. Constraint-based filtering with a greedy information-gain heuristic. 85%+ solve rate within six guesses across a 5,922-word list, covered by a 36-test suite. | Python Pillow |
Agentic AI & LLM
Primary languages
Backend & Infrastructure
Frontend
Tools
Practices — Context Engineering · Prompt Engineering · Evaluation Design · Behavioural Testing · Red Teaming · Chain of Density · Tree of Thoughts · Token Sampling
Also used — C/C++, C#, Java and .NET/ASP.NET in university coursework. Listed for completeness, not claimed as working skills.
Stats images generated by jstrieb/github-stats.
AI Automation Engineer (Intern), HubIT — June to July 2026, Karachi. Built a Python pipeline extracting recurring utility documents into schema-validated billing records for a residential property management client, with repair-and-retry on malformed model output so bad records were blocked before billing. Designed the reconciliation stage matching records to individual units, hardened against silent mis-mapping with structured failure logging.
Engineering Intern, Pakistan Aeronautical Complex (PAC) Kamra — Jan to Feb 2024. Analysed systems integration workflows, avionics production pipelines and precision manufacturing across the JF-17 Thunder programme.
- Extending Rahzaan so any student in Karachi can use it
- Independent research on LLM behavioural failure modes: context anchoring and long-horizon drift
- Preparing Masters applications for AI and intelligent systems programmes, Winter 2027/28
