Senior AI Research Engineer

About Us At SkyLabs AI Inc.

, we are at the forefront of the artificial intelligence revolution.

As a US-headquartered company, we conduct applied research on AI for intelligent reasoning.

We specialize in complex neurosymbolic AI to solve intricate problems within software engineering.

Our team is composed of world-class researchers and engineers dedicated to building the platforms and intelligen...t agents that will power the next generation of software.

If you are passionate about building truly intelligent systems and want to make a lasting impact, join us.

The Role We are seeking an exceptional Senior AI Research Engineer with a strong focus on training and improving LLMs across the full lifecycle —from Domain‑Adaptive Pretraining to Supervised Fine‑Tuning (SFT) , RL/RLVR , and advanced post‑training techniques (e.

g., reward modeling, preference optimization, RL‑VR style workflows).

You will lead hands‑on training efforts, contribute to research direction, and build robust pipelines for data, training, evaluation, and deployment.

You should have in‑depth understanding of LLM internals (attention/MLP dynamics, normalization, optimization behavior, scaling laws), and modern architectures including Mixture of Experts (MoE) and other frontier design choices.

You will work closely with product and platform teams to translate research ideas into performant systems.

Salaries in USD (income tax exemptions) Work in Pakistan Timezone Comprehensive health allowance Monthly team events and activities Relocation allowance (if you're moving to Islamabad) Opportunity to work with top minds in the industry and academia Startup culture where ideas are heard at all levels Adequate annual, sick, casual and parental leaves Key Responsibilities Lead the team of AI Engineers/Researchers Train and iterate on LLMs end-to-end across DAPT/CPT, SFT, and RL-based post-training (preference optimization, reward modeling, verifiable rewards, policy optimization, and related variants).

Design training recipes : tokenization strategy, curriculum/sampling, batch/sequence packing, optimizer + scheduler choices, stability techniques, and hyperparameter search.

Architect and implement efficient training systems using distributed training (data/tensor/pipeline parallelism), FSDP/ZeRO, activation checkpointing, mixed precision, and throughput optimization.

Develop and maintain LLM data pipelines : large-scale data ingestion, filtering, deduplication, contamination checks, domain balancing, safety filtering, and dataset versioning.

Perform deep EDA and root-cause analysis for training/eval regressions (loss spikes, instability, alignment drift, memorization, toxicity, domain overfitting), and implement mitigations.

Build evaluation and benchmarking suites : offline metrics, human preference pipelines, automatic evals, domain-specific test sets, adversarial probing, and model behavior tracking over time.

Apply and advance model compression and efficiency : distillation from LLMs, quantization-aware considerations, pruning/sparsity (including MoE routing behavior), and inference-time optimization tradeoffs.

Implement and operationalize LLMOps : experiment tracking, reproducible runs, model registry, training telemetry, artifact management, CI for training configs, and safe rollout strategies.

Leverage AI coding tools (e.

g., Cursor, Copilot) to accelerate development while maintaining strong engineering rigor, testability, and code review standards.

Collaborate cross-functionally with platform, infra, and product teams to integrate trained models into agentic/production pipelines (retrieval, tool use, evaluation gates, monitoring).

Stay current with the field and translate new ideas (architectures, post-training, data strategies, evals) into internal experiments and improvements.

Qualifications & Skills 5+ years in ML engineering / applied research, with significant hands-on experience training large models , especially LLMs.

Strong practical knowledge of transformer internals and modern variations (MoE, routing, load balancing, attention variants, normalization choices, long-context strategies).

Demonstrated experience with DAPT/CPT , SFT , and RL-based post-training (e.

g., reward modeling, preference datasets, verifiable rewards, policy optimization).

Ability to reason about why a method works and when it fails.

Excellent understanding of data science fundamentals : EDA, dataset design, bias/contamination, sampling strategies, and error analysis.

Proven ability to root-cause and debug training instabilities and performance regressions (numerics, data issues, infra bottlenecks, config drift).

Expert-level programming in Python , with strong software engineering fundamentals (data structures, algorithms, systems thinking).

Experience with key tooling such as PyTorch , distributed training frameworks (FSDP/DeepSpeed/ZeRO), and training orchestration.

Familiarity with LLMOps : tracking (e.

g., W&B/MLflow), dataset versioning, reproducibility, deployment constraints, and monitoring of model behavior.

Comfortable building agentic pipelines around LLMs (evaluation agents, data generation/labeling loops, model-in-the-loop distillation, automated red-teaming).

Bonus: publications, open-source contributions, or demonstrated leadership in shipping LLM training improvements.

Who You Are You’re a pragmatic researcher-engineer: you can read papers, run experiments, and also build the training and data systems that make results reproducible and shippable.

You’re meticulous about data and measurement, fast at debugging, and comfortable making tradeoffs between quality, cost, and timeline.

You communicate clearly, mentor others, and thrive in an environment where you’re trusted to drive big initiatives.

Show full description