SecNinjaz Technologies LLP

Openings: 1 | Experience: around 2 years

Location / work mode: Delhi

About the role

Build the data, evaluation and model-improvement capability behind AI-powered cybersecurity systems.

Turn reviewed security work into trustworthy datasets, measure performance on unseen cases, and run reproducible language-model post-training experiments.

You will develop and test Python pipelines, working with cybersecurity specialists and an agent/runtime engineer. You need strong applied AI/ML engineering and practical cybersecurity ability to judge evidence, understand labels and determine whether an experiment improves a security workflow.

What you will do

- Turn approved workflow traces, tool outputs, failures, expert corrections and verified outcomes into versioned datasets, with provenance, permitted-use tracking, sensitive-data handling, deduplication and separate training, validation and held-out test cases, grouping related applications, engagements and near-duplicate traces to prevent leakage between splits.
- Work with security specialists to define tasks, labels, negative cases and independently checked outcomes; investigate ambiguous labels, false positives and data leakage.
- Build repeatable evaluations for task success, finding quality, tool use, regressions, latency and cost.

Compare model and harness changes against versioned baselines, isolate changes where practical,

and document interactions, uncertainty and evaluation limitations.

- Run reproducible language-model post-training experiments using supervised fine-tuning,

preference-based methods or reinforcement learning where appropriate; explain the chosen method and investigate misleading reward signals.
- Jointly define versioned trace and tool interfaces, evaluation criteria and release checks with the agent/runtime engineer and cybersecurity specialists. Security specialists review domain labels and validate findings.




- Work with the runtime engineer on controlled execution environments and feedback for evaluation and training experiments. Keep operational memory, training data and held-out evaluation cases distinct.
- Package datasets, configurations, model checkpoints and results so another engineer can reproduce the work; document failures and recommend whether an improvement warrants adoption.
- Integrate accepted models into the agent workflow with compatibility tests, regression checks and versioned rollout/rollback, including private or self-hosted deployment where required.

SecNinjaz Technologies LLP 1What you should bring

- Around two years of hands-on development experience; strong Python, data processing, Git, debugging and automated testing.
- Practical AI/ML knowledge: training versus inference, loss functions, overfitting, generalisation, data quality, metrics and experimental comparison; understanding of LLM behaviour, prompting, retrieval and tool use.
- A reproducible language-model fine-tuning or post-training project covering data preparation, a baseline, held-out results and failure analysis. A well-executed prototype or implemented coursework qualifies; be ready to explain and modify your implementation.
- Working knowledge of dataset provenance, label review, deduplication and leakage prevention.
- Reinforcement-learning fundamentals: environments, actions, rewards, evaluation and the risk of rewarding the wrong behaviour.
- Demonstrated practical cybersecurity ability: reason about authentication, authorization, trust boundaries and common web/API or code weaknesses; interpret tool output, reproduce and validate a finding, reject a false positive and verify a fix in a controlled environment.




- Ability to integrate an LLM into a Python application or agent workflow, work with APIs and Linux tooling,

and explain experimental results clearly.

Relevant skilled work, labs, research and personal projects can demonstrate these skills.

Good to have

- Preference optimisation, RL fine-tuning, distillation or comparison of post-training methods.
- Experiment tracking, dataset versioning and training or inference performance work.
- Open-weight model serving, quantisation or private deployment.
- Deeper security research, detection engineering or security evaluation experience.

Initial outcomes, with the team

- Establish a permitted pipeline from workflow traces to reviewed, versioned datasets and an evaluation baseline.
- Run a bounded post-training experiment and compare it with the base model on held-out security tasks,

documenting limitations and failures.
- Demonstrate integration of the candidate model in the agent workflow and recommend adoption or rejection using quality, reliability and cost evidence.

Selection process A discussion of an existing post-training project and a short practical exercise using synthetic or sanitized security data. You will interpret evidence, review labels and permitted use, identify leakage risks,

implement a small data or evaluation improvement, and explain a baseline comparison and adoption decision.

The exercise does not require a large new training run. AI tools may be used; explain your contribution and be ready to debug your solution. Nothing is run against real systems.

Apply

Send your CV and a relevant fine-tuning/post-training project link, or a short technical write-up covering your contribution, data preparation and evaluation, to [application contact/link]. Use "AI/ML Engineer

— Data, Evaluation & Model Improvement" as the application subject.

All security work at SecNinjaz is authorized and operates under agreed rules of engagement.

📌 AI/ML Engineer — Data, Evaluation & Model Improvement (Vijayawada)
🏢 SecNinjaz Technologies
📍 Vijayawada