Junaid Anwar Qader
Toronto · Graduating Dec 2026 · Open to AI engineering roles

I build the systems between an AI model and a dependable product.

Applied AI engineer working on agentic systems, governed knowledge, retrieval evaluation, and multimodal machine learning. I turn research claims into measurable, production-shaped systems.

Résumé Email me
LinkedIn GitHub
01 / SELECTED WORK

Systems with an argument behind them.

The problem, the engineering choice, and the evidence—not just a stack of tools.

Enterprise agent infrastructure2026

Skills Registry & Collective Intelligence

At Horizon3 AI Labs, I am building the infrastructure that lets enterprise agents use organizational knowledge without turning it into an untraceable prompt.

The Skills Registry converts documents into reviewed, versioned procedures, retrieves only the relevant ones for an agent, and preserves the source and version behind every delivery. The Collective Intelligence work closes the loop by capturing agent transcripts across frameworks, finding reusable knowledge in those interactions, and routing it through the same governance process.

The design treats memory as a security boundary: new memories are quarantined before approval, retrieval is scoped to the right user and project, and forgetting removes data from the serving path immediately.

FastAPIMCPGoogle ADKSQLAlchemyIR evaluation
Ground truth, degraded input, teacher output, and student output comparisons for under-display camera restoration
Computer vision · course research2025

Frequency-Aware Distillation for UDC Restoration

Transferred the spectral reasoning of a MambaIR teacher into a lightweight U-Net with amplitude and multi-scale phase losses. Full-image inference fell from about 2.59 seconds to 61–93 milliseconds while preserving teacher-level perceptual quality.

Code & report ↗~28–42× faster
PyTorchMambaIRFFT lossesU-Net
Confusion matrices comparing multimodal Bi-LSTM Transformer and ensemble approaches for three ICU length-of-stay classes
Clinical machine learning · SickKids2026

Pediatric ICU Length-of-Stay Prediction

Built a leakage-safe pipeline from the first six hours of ECG, PPG, and ABP waveforms plus EHR data for 142 pediatric liver-transplant patients. I owned beat validation and Windkessel-based ABP feature extraction, then compared tabular, ensemble, and Bi-LSTM + Transformer approaches under nested cross-validation.

Macro-F1.553 → .573
Macro-AUPRC.633 → .674
EHR-only baseline → best waveform model

What the result actually says: the waveform pipeline was technically viable and directionally better, but overlapping 95% confidence intervals meant the improvement was not statistically significant. The limiting factor was a small, single-centre cohort—not a number worth overselling.

PyTorchtime seriessignal processingclinical ML
02 / EXPERIENCE

Research depth, production instincts.

2026

Applied Research Intern (MScAC)

Unilever · Horizon3 AI Labs · Toronto

  • Architected a versioned Skills Registry with governed ingestion, human review, rollback, deterministic context assembly, and REST/MCP delivery.
  • Built research harnesses to compare skill generation and retrieval methods before they become product decisions.
  • Developed a framework-agnostic transcript SDK and collector with reliable background delivery, scoped identity, and line-level citations.
2026

Machine Learning Researcher

The Hospital for Sick Children · Toronto

Built a leakage-safe ICU length-of-stay pipeline from six hours of high-frequency waveforms plus EHR data for 142 pediatric liver-transplant patients. Owned Windkessel-based ABP feature extraction and compared tabular, ensemble, and Bi-LSTM + Transformer approaches under nested cross-validation; the best waveform model improved both macro-F1 and macro-AUPRC, though overlapping confidence intervals made the result directional rather than significant.

2024

Software Development Engineer Intern

Wells Fargo · Chennai

Built secure developer-assistant extensions for VS Code and IntelliJ: retrieval over Confluence plus Harness build logs and pipeline state in the editor, using JavaScript, Java/AWT Swing, NLTK, and spaCy.

2024–25

Undergraduate Researcher

National Institute of Technology, Warangal

Developed deep-learning systems for Cathepsin inhibitor screening and acoustic mosquito detection. The work led to first- and third-author peer-reviewed journal publications.

03 / SELECTED PROJECTS

A short, relevant project list.

The work most useful for AI and software engineering conversations. Everything else remains available on GitHub.

012026

Skills Registry & Collective Intelligence

Governed organizational knowledge, context delivery, transcript capture, and agent memory.

FastAPIMCPagents
Private internship work
032025

SurgiSync ↗

Voice-driven agent for ER admissions, transfers, discharges, and patient queries.

agentsspeechtool calling
042025

Text-to-SQL Agent ↗

Schema-aware LangGraph pipeline with SQL execution and error correction.

LangGraphSQLiteLlama 3
072024–25

CathepsinDL ↗

Deep-learning screening of Cathepsin inhibitor potency from molecular descriptors.

drug discovery1D CNNpublished
082024

CaptionCrafter ↗

Deployed CNN-LSTM image captioner with an LLM layer for social-ready copy.

TensorFlowStreamlitGroq
More experiments and coursework are available without crowding this page.Browse GitHub ↗
04 / RESEARCH

Evidence is part of the build.

Peer-reviewed publications

Languages

Python · Java · C++ · SQL · JavaScript · Bash

AI & ML

PyTorch · scikit-learn · XGBoost · Transformers · CNNs · LSTMs · signal processing

Agents

LangGraph · LangChain · Google ADK · MCP · RAG · embeddings · LLM evaluation

Systems

FastAPI · REST · SQLAlchemy · Alembic · Docker · AWS · MLflow · GitHub Actions · pytest

LET'S BUILD SOMETHING THAT HOLDS UP

Available for AI engineering roles.

Toronto-based, graduating December 2026. Open to roles in applied AI, ML engineering, agent infrastructure, retrieval, and multimodal machine learning.

junaid24@cs.toronto.eduLinkedIn ↗