model: purvi-verma
type:  applied-ai-engineer · 2+ yrs production LLM systems

Applied AI engineer who spent two years shipping LLM safety systems, multi-agent pipelines, and multimodal CV+LLM tooling into production at enterprise scale. Left the role to re-train from checkpoint — currently an M.Tech candidate in AI at IIT Roorkee.

base_model:      B.Tech, IIT Bhubaneswar (CGPA 8.2/10)
fine-tuned_at:   Publicis Sapient — Senior Associate, Data Science (Jul 2024 – Jul 2026)
pre-training:    AI Intern, Blockatena (Mar 2023 – May 2023)
currently:       M.Tech, Artificial Intelligence — IIT Roorkee (2026 – 2028)
stack:           LangChain · LangGraph · RAG · MCP · vLLM · LoRA/PEFT
status:          in production for 2+ years, now training on harder problems

Bodhi Compliance — LLM safety guardrails Enterprise LLM deployments were producing non-compliant outputs at scale. Architected safety guardrails that cut harmful output rates from 45% → under 5%.

Multi-turn agent orchestration Single-turn LLM interactions couldn't handle complex enterprise workflows. Designed a context-aware, tool-calling agent that raised task success rate from 70% → 92% and enabled systematic A/B testing across model variants.

Bodhi Vision — multimodal CV + LLM pipeline Document processing for enterprise digital assets was fully manual. Built an end-to-end pipeline (format conversion, vLLM-powered segmentation, text detection) hitting 90.3% extraction accuracy, 87% semantic consistency, and zero failures across 500+ assets — replacing manual review entirely.

Eli Lilly — regulated pharma content at scale Scaled the CV + LLM pipeline into a compliance-sensitive pharmaceutical environment, automating processing of 10,000+ assets with high accuracy in a domain where errors aren't an option.

flowchart LR
    A["B.Tech, IIT Bhubaneswar"] --> B["AI Intern<br>Blockatena"]
    B --> C["Senior Associate<br>Data Science, Publicis Sapient"]
    C --> D["Bodhi Compliance<br>45% → 5% harmful outputs"]
    C --> E["Multi-turn Agents<br>70% → 92% task success"]
    C --> F["Bodhi Vision<br>90.3% extraction accuracy"]
    F --> G["Eli Lilly<br>10,000+ assets automated"]
    C --> H{"decision layer"}
    H -->|left to go deeper| I["M.Tech AI<br>IIT Roorkee"]

    style H fill:#0d1117,stroke:#4a9eff,color:#4a9eff
    style I fill:#0d1117,stroke:#4a9eff,color:#4a9eff

safeguard_toolkit — pip install safeguard_toolkit A modular Python security scanner with 4 composable scanners for hardcoded secrets, risky configs, vulnerable dependencies, and unsafe file permissions. Entropy-based secret detection, dependency parsing across requirements.txt/Pipfile/pyproject.toml, CI/CD-ready plugin support.

CodeMap — AI-powered code architecture analyzer A 7-stage pipeline combining static analysis with LLM reasoning to auto-generate architecture diagrams, dependency graphs, and docs from any repo. Batched summarization (10 functions/call), automated design-pattern detection, and self-healing Mermaid diagrams — an LLM repairs invalid syntax before render. Full analysis in under 5 minutes.

contributions — google/bespoke (error handling & API key resolution) · video-keyframe-detector (structured logging, core bug fixes) · CRAFT-pytorch (PyTorch/torchvision compatibility, CUDA/CPU fallback)

$ curl -X POST https://api.purvi-verma.dev/v1/connect \
    -H "Content-Type: application/json" \
    -d '{"intent": "collaborate"}'

{
  "response": "always open to it",
  "endpoints": {
    "linkedin":  "https://www.linkedin.com/in/purvi-verma-3a553a23b/",
    "portfolio": "https://purvi1508.github.io/PurviVerma/",
    "leetcode":  "https://leetcode.com/u/20ce01050/",
    "email":     "purviverma2026@gmail.com"
  }
}