Projects

Built a zero-trust proxy that sits between LLM agents and MCP tool servers, enforcing policy on every tool call instead of trusting the model with direct access — no code changes required to the agent or the tools it calls.

Implemented an OPA/Rego policy engine combined with taint tracking and LLM-based semantic risk scoring, so requests are evaluated for intent and data flow, not just static permissions — with human-in-the-loop approval workflows for high-risk actions.

Hardened through four rounds of adversarial security audits, improving the platform's security posture from 2.1/10 to 9.1/10. Selected for the Claude for Startups program, the Zendesk Startup Program, and the Sarvam AI Startup Program. Live at runwall.in.

An ultra-fast NLP preprocessing library that combines tokenization, spell correction, and parallel text cleaning into one package — replacing the usual NLTK + spaCy + BeautifulSoup + TextBlob stack with a single import.

Tokenization recognizes complex patterns as single tokens (currency like `$20` or `₹100`, emails, hashtags, mentions, phone numbers, URLs, dates, emojis, contractions, hyphenated words), and LightningSpellCorrector handles noisy input alongside configurable HTML/URL/emoji removal.

Ships with parallel batch processing via `batch_preprocess` — the README reports ~3 minutes for 1M documents versus 45 minutes for NLTK, on 0.8 GB of memory. Published on PyPI and MIT-licensed.

Orchestrated ingestion of speech transcripts, on-screen OCR, and creator metadata so each video becomes a structured bundle LangGraph stages can reason over with deterministic fallbacks.

Grounded compliance reasoning with Azure AI Search and curated policy snippets—reducing “creative hallucination” when judging claims, disclaimers, or risky visuals.

Exposed reviewer dashboards via FastAPI + Next.js so legal and marketing teams can replay decisions, override labels, and export audit trails for regulators.

Versioned data and pipeline artifacts with DVC and tracked experiments in MLflow (with a staging/production model registry on DagsHub) so every training run is reproducible and rollbacks are a registry transition, not a code change.

Packaged the trained model in a Docker image, pushed to AWS ECR, and deployed to Amazon EKS (provisioned with eksctl and CloudFormation) — exposing it through a Flask service behind Kubernetes Deployment and Service manifests.

Wired GitHub Actions for CI/CD on every push and stood up Prometheus + Grafana on dedicated AWS EC2 nodes to scrape the cluster — so latency, error rate, and request volume are visible the moment a new revision rolls out.

Built an end-to-end pipeline that ingests camera and sensor-derived signals, applies graph-style modeling (GNN) on behavioral relationships, and surfaces anomalies ranked by severity instead of naive threshold alarms.

Integrated vector search (Pinecone) with Gen-AI summarization so investigators get short, grounded narratives tied to similar past incidents rather than raw clip dumps.

Delivered multi-channel alerting (WhatsApp, email, phone) with escalation policies so on-call teams see concise context within seconds of a confirmed trigger.