# Ansuman SS Bhujabala

> AI Engineer working on agentic systems, retrieval-augmented generation, and LLM evaluation. Based in Hyderabad, India. Open to AI Engineer, AI Architect, AI Safety Engineer, and generalist roles — remote or relocating.

## Profile

- Role: AI Engineer (CTO)
- Focus: agentic systems, RAG, LLM evaluation, adversarial testing, Model Context Protocol, AI security, AIOps
- Location: Hyderabad, India
- Contact: ansumanbhujabal1@gmail.com
- Site: https://ansumanbhujabal.github.io
- GitHub: https://github.com/Ansumanbhujabal
- LinkedIn: https://www.linkedin.com/in/ansuman-simanta-sekhar-bhujabala
- Résumé: https://ansumanbhujabal.github.io/resume.pdf

## Experience

### AI Engineer (CTO) — Anyfeast · Feb 2025 – Present · London, UK (Remote)

- Leading AI engineering for a real-time personalised meal-kit AI ecosystem spanning 50+ dietary verticals including complex health constraints, with validation and evals.
- Architected and scaled a complete harness-engineered recipe engine with guardrails and validation layers, producing 130,000+ diverse recipes serving thousands of customers across the UK and India.
- Led an 8-member team (engineering and doctors) to build an agentic consultation system grounded in verified nutrition data.

### AI Engineer Intern — Invest4Edu · May 2024 – Jan 2025 · Hyderabad, IN

- Designed ETL pipelines for educational data scraping handling 10M+ records.
- Built a RAG-based Q&A system over scraped educational content using open-source LLM fine-tuning; added in-house evals and observability with QA teams, achieving 78% accuracy with refusal handling.

## Selected work

### Polly Harness (live)

A nine-layer refund agent built specifically to be attacked. LangGraph, Qdrant, custom MCP tools.

Metrics: 88.6% pass across 229 adversarial cases; 98.7% injection resistance; 100% jailbreak resistance

Repository: https://github.com/Ansumanbhujabal/Polly_Harness | Demo: https://ansumanbhujabal-polly-harness.hf.space

### UltraDoc Intelligence (live)

Hybrid retrieval over logistics and TMS documents — dense plus BM25 with reciprocal rank fusion, four deterministic guardrail layers, and an LLM-as-judge evaluation framework, traced in Langfuse.

Metrics: 91.9% answer accuracy; $0.003 per query

Repository: https://github.com/Ansumanbhujabal/ultradoc-intelligence- | Demo: https://ansumanbhujabal-ultradoc-intelligence.hf.space

### Mentis (live)

Procurement intelligence briefs for medical substances — product profile, clinical use, market, manufacturers, regulatory status, sourcing and risks, with every claim backed by a verified source URL. Four-step safety escalation, versioned prompts, per-call cost accounting, observable traces.

Repository: https://github.com/Ansumanbhujabal/Mentis | Demo: https://huggingface.co/spaces/AnsumanBhujabal/Mentis

### Pugmark (live)

Turns hunting and natural-history novels into illustrated, evaluated bestiaries. Five-stage async pipeline — ingest, extract, validate, enrich, render — with LiteLLM provider abstraction (Gemini primary, Groq fallback, Ollama offline), Wikidata validation, and AI-illustrated PDF export.

Repository: https://github.com/Ansumanbhujabal/Pugmarks

### Pulse (early access)

End-to-end encrypted messaging. X25519 pairing ceremony, HKDF-derived subkeys, AES-256-GCM. Security review found the six-emoji verification phrase grindable at 2^36 by the relay server — precisely the adversary the design distrusts. Mitigated with iterated hashing.

Demo: https://pulseldr.onrender.com/