I’m obsessed with building the best agents, extracting every ounce of accuracy under realistic latency and token budgets, backed by extensive trace evals.
AboutTurning research on efficient agents into products people can rely on.
Experience
Hover a row for the detail.Click a row for the detail.
Engineering
Résumé ↓- 01Brainbox AutomationsAI EngineerDec 2025 – Now
- Deployed a safety-guardrailed AI coaching backend on GCP Cloud Run (FastAPI, Vertex AI) serving 300+ minor and adult athletes
- Cut CI/CD deploys from ~2 min to 10 s with GitHub Actions, keeping an SSH fallback
- Built SQL-aware agents over 50+ Supabase tables, answering 50% faster
- Shipped chat and VAPI voice AI for Earlibird (AU) handling 200+ chats a day and 5,000+ calls, with 80% resolved by AI and 21% booking conversion
- Summarised hour-long sales calls with a parallel map-reduce over Gemini on AWS Lambda
- 02Epistemy UKSoftware Engineer (AI) & Team LeadSep – Dec 2025
- Led the end-to-end build of an AI tutoring platform with a Nest.js backend (20+ endpoints) and a Next.js frontend
- Orchestrated multi-agent workflows on an event-driven Redis/BullMQ queue with fault tolerance
- Held 90%+ unit-test coverage with CI and pre-commit hooks, leading a team of 2
- 03CogniMind AIMachine Learning InternFeb – Apr 2025
- Raised VLM extraction accuracy by 20% through prompt engineering
- Sped up retrieval and inference by 10% with quantization and HNSW
- Automated MLOps with Dockerized Airflow and 5+ DAGs, cutting manual work by 60%
- Built Docker and GitHub Actions CI/CD that cut deploy time by 30%
- 04RapidsAI Machine Learning InternSep – Dec 2024
- Built a RAG chatbot (Streamlit, FastAPI) with contextual session management
- Reduced errors by 50% with chain-of-thought prompting
- Halved OpenAI API costs (−50%) with a complexity-based multi-model router
Research
Academic CV ↓- 01IPT Lab, NUSTResearcherJun 2026 – Now
- Certified small-to-large escalation for tool calls, keeping 67% of requests on the small model at a 30% error target, where tuned thresholds broke it about half the time
- Built agents that repair function-calling benchmark keys. Over 1 in 10 BFCL Live and ACEBench items reject a correct call, and the fixes lift a 32B model by up to 12 points
- Certified early exit in a wireless foundation model, saving 14–39% of encoder compute at 95% confidence
- Showed the guarantee breaks when SNR is estimated, and restored it by certifying on estimated groups
- Co-authored 2 studies on what encryption fails to hide, where agent service discovery reveals a drone’s mission at 0.947 balanced accuracy against 0.25 chance
- 02Bradbury LabResearch Intern · remoteApr 2025 – Jan 2026
- Proposed a training-free layer-merging method based on Tucker decomposition to cut parameter count
- Analysed self-attention to test aligning Query and Key projections in efficient-by-design architectures
- Reviewed Transformer topology and parameter-efficient fine-tuning, focusing on weight sharing
- 03MachVis LabUndergraduate Research InternOct – Dec 2025
- Built an LLM-based factual-verification framework for the clinical accuracy of generated pathology reports, beyond BLEU/ROUGE
- Curated a challenge set of 100+ gigapixel whole-slide images with real artifacts and stain variation
- 04NUST · remoteUndergraduate Research InternJun – Sep 2025
- Combined MedSAM with meta-learning for few-shot dental radiograph segmentation, +12% on scarce disease classes
Inside the research
Agentic AI under resource constraints. Long-horizon agents that plan, call tools and manage context on small models, where every call has a cost, and making those choices reliable rather than hand-tuned.
A hundred requests ranked by confidence. The 67 small dots stay on the small model and the large blocks escalate, under a certified 30% error target that tuned thresholds miss about half the time.
A large teacher distilled through a narrow channel into a small task-specific student.
Over 1 in 10 BFCL Live and ACEBench items reject a correct call. Agents repair the broken keys, and a 32B model scores up to 12 points higher.
Encrypted, yet a drone’s service discovery still gives away its mission, at 0.947 balanced accuracy against 0.25 chance.
- Of tool calls kept on the small model at a certified 30% error target
- 67%
- Accuracy points a 32B model regains once broken benchmark keys are repaired
- +12
- Balanced accuracy at reading a drone’s mission from encrypted service discovery
- 0.947
- Of the DeepSeek-VL2 (MoE) teacher’s performance retained by LiteDoc, on average
- 89.6%
- C1DOI ↗ (opens in a new tab)
LiteDoc: Distilling Large Document Models into Efficient Task-Specific Encoders
Raza, Taha, Ulges, Schwanecke, Moetesum, Shafait, ICDAR 2026, Springer LNCS
- S1Under review
ESCORT: Escalation with Statistical Certification of Risk Targets for Language-Model Service Brokers
Taha, Zeeshan, Hassan, Cirillo, Ahmed, Hassan, IEEE Transactions on Services Computing, 2026
- S2Under review
AMEND: Agent-Mediated Errata for Noisy Datasets of Function-Calling Benchmarks
Taha, Zeeshan, Ahmad, Hassan, IEEE ICC 2027
- S3Under review
Flight Plans in Plain Sight: Intent Leakage in Agentic Service Discovery for Low-Altitude Wireless Networks
Ahmad, Zeeshan, Taha, Dev, Jung, Hassan, IEEE Vehicular Technology Magazine, 2026
- S4Under review
Observable Participation: A Passive Aerial Threat to Secure Aggregation in HAPS Federated Learning
Ahmad, Zeeshan, Taha, Hassan, IEEE ICC 2027
- P1In prep.
Early Exit in a Wireless Foundation Model When the Signal-to-Noise Ratio Must Be Estimated
Taha, et al., IEEE Transactions on Machine Learning in Communications and Networking, 2026
Selected projects
- { 01 / 04 }
create-rag-app
GitHub ↗ (opens in a new tab)A Python CLI that scaffolds production-ready, Dockerized RAG apps from 13 Jinja2 templates, with 3 retrieval strategies and 8+ integrations.
- { 02 / 04 }
Bare-metal Prospector
GitHub ↗ (opens in a new tab)A framework-free SDR agent on a raw Python ReAct loop, with a syntax-directed parser that heals tool hallucinations and a three-tier memory.
- { 03 / 04 }
GPT-2 from scratch
GitHub ↗ (opens in a new tab)Pretrained GPT-2 from scratch in PyTorch, sped up inference with KV caching and speculative decoding, and instruction-tuned it on Alpaca.
- { 04 / 04 }
Agentic RAG
GitHub ↗ (opens in a new tab)A LangGraph planner and re-planner agent for retrieval, evaluated with RAGAS at over 95% faithfulness.