Doctors tell us things they would never post publicly: a patient's case, a mistake on a night shift, how burnt out they really are. That trust is the most valuable thing Recollia has. We need one person whose job is to protect it.

Recollia builds AI for medical students and doctors: tutors, clinical cases, a memory layer that remembers each student, a companion for the heavy weeks, and a media engine. You'll own security and privacy across all of it, from what we run today to everything on our roadmap.

WHERE THIS GOES

This starts as an internship with a stipend of Rs 20,000 and a possible incentive of Rs 5,000 every two weeks, depending on performance. Make Recollia the AI product doctors trust with their most sensitive words, and it converts into a founding engineer role with equity. The bar is high on purpose: in healthcare, one leak can end a company.

WHAT YOU'LL BUILD AND OWN The de-identification layer

- A layer every input passes through: detect and redact patient names, phone numbers, IDs, addresses, dates and other identifiers before anything reaches a model, a log, an analytics event or a training set
- Built for Indian medical text: Indian names, Aadhaar and phone formats, hospital shorthand, Hinglish and regional languages
- Measured, not assumed: recall on a labelled test set, because one missed patient name is a breach
- Reversible only where it must be, with keys only the right systems can use

AI and LLM security

- Prompt injection and data leakage across Claude, OpenAI and our own models
- What we send to model providers, and on what retention terms
- Memory and connectors: what our memory layer is allowed to remember, for how long, and who can ever read it
- Training data hygiene:



nothing sensitive goes into a fine-tune

Cloud and application security

- AWS done right: IAM least privilege, VPC design, encryption at rest and in transit, secrets management, CloudTrail, GuardDuty
- Application security: authentication, authorization, OWASP Top 10, dependency and container scanning in CI
- Penetration testing our own product before anyone else does

Privacy and compliance

- India's DPDP Act: consent, purpose limits, data principal rights, retention, breach notification
- The policies, data maps and audit trails that let us answer "where is this doctor's data, and who touched it?" in minutes
- Ready for the security questionnaires hospitals and colleges will send us

Security by design

- Threat-model every new feature before it's built, not after it ships
- Incident response: the runbook, the alerts, and the calm head when something goes wrong
- Make the secure way the easy way for the rest of the team

WHAT YOU NEED

Security fundamentals

- You've secured a real system that handled sensitive data, and you can explain what you'd attack first
- Threat modeling, OWASP, authentication and authorization, cryptography you use correctly rather than invent
- Offensive instinct: you've broken things on purpose (CTFs, bug bounties, pentests)

Privacy engineering

- PII detection and redaction: NER models, rules, tools like Microsoft Presidio,



and their failure modes
- De-identification versus anonymization, and why the difference matters in medicine
- DPDP Act familiarity; HIPAA or GDPR experience is a plus

AI security

- Prompt injection, jailbreaks, data exfiltration through tools and connectors, and how to contain them
- How LLM providers handle data, and how to design so it doesn't matter

Cloud and engineering

- AWS security in depth; infrastructure as code (Terraform or CDK)
- Python or TypeScript well enough to build the de-identification layer yourself, with tests
- Logging, monitoring and alerting that catch the real thing without drowning the team

No degree required.

Skills over credentials.

Certifications are welcome but don't replace proof.

WHAT HUNGER LOOKS LIKE TO US

- You read breach reports for fun and think about how you'd have stopped them.
- You'd rather find the hole yourself than learn about it from a doctor.
- You can say no to a feature and offer the secure way to build it in the same sentence.
- You ask for more scope, not more instructions.
- You say when something is broken, including our decisions.

HOW TO APPLY

Email [email protected] with the subject "Security & Privacy" and:
- One thing you secured, broke or built: a repo, write-up, CVE, bug bounty or CTF result
- In three sentences: what it was, what the weakest point was, and what you did about it
- A doctor types this into an AI tutor: "My patient Ramesh Kumar, 54, from Madurai, phone 98xxxxxx21, came in with chest pain yesterday." In five sentences or fewer, tell us everything that should happen to that message before it reaches a model, and what most systems get wrong.

📌 AI Security & Privacy Engineer (India)
🏢 Recollia
📍 India