AI Engineer. I build LLM applications, AI agents and RAG systems that hold up on real, messy data, and I measure them so you know they work.

Lead AI Engineer at Effulgent (member of the OpenAI, Anthropic and Google Cloud partner programs). M.S. in Applied Data Science, Clarkson University. 6+ years of Python and software engineering.

• AI agents and agentic workflows: LangGraph, tool/function calling, MCP servers, human-in-the-loop approvals
• RAG and document intelligence: hybrid search (vectors + full text), reranking, answers that cite their sources
• LLM integrations: Claude, OpenAI and Vertex AI with structured outputs, guardrails and cost controls
• Evaluation: test sets, LLM-as-judge, regression checks, so changes are measured instead of guessed
• Production plumbing: FastAPI, Postgres/pgvector, job queues, Docker, CI/CD with rollback, monitoring

Turns résumés and job notes into evidence-backed shortlists for an A/E/C staffing workflow, on top of an existing ATS.

• LangGraph ingest per résumé: dedupe → parse and screen → Claude structured extraction → code validation → Voyage embeddings, with a Postgres job queue and workers
• Hard filters + hybrid retrieval (pgvector + Postgres full text) + rerank: recall@10 48% → 89% at 98% precision@10
• Evidence-backed match scoring, where every claim is checked against a quoted résumé line: 97.6% submit precision, about $0.02 per résumé
• Evaluated on a 597-résumé test set, including scanned, corrupt, duplicate and prompt-injection files
• Deployed on Google Cloud with Docker, GitHub Actions auto-deploy and rollback, LLM spend caps and LangSmith tracing

Python LangGraph LangChain Claude Voyage AI pgvector FastAPI Next.js n8n Docker GCP

An MCP server that lets Claude or any MCP client search, quote and compare Kalshi and Polymarket markets: order books, recent trades, price history, and cross-venue pairs ranked by price gap.

• 6 read-only tools on one venue-neutral data model (every price is a YES probability; NO-side quotes and fills converted)
• Built for agent traffic: per-venue rate limiting, retries that honor Retry-After, a de-duplicating TTL cache, inverted-index search
• Cross-venue matching that rejects look-alikes (different thresholds, above vs touch, cut vs hike, negation), with regression tests from real title pairs
• 12 regression-tested look-alike rules; 0.3 s search, zero rate-limit errors; a live cross-venue board rebuilt every 6 hours

Python MCP httpx asyncio pydantic

Each one is a public repo with Docker, CI, design decisions and an evaluation on a public benchmark, with every number from a real run.

PDF → text → custom spaCy NER that extracts sales, address and tax-parcel data from legal documents (92% accuracy on 500+ documents), with per-entity precision/recall/F1 evaluation. Built during my graduate research at Clarkson.

Python spaCy pdfminer NLP

Entry in IMC's global trading competition: order-book strategies for three products (fair-value market making, a price-forecast strategy and position-capped sizing), using only numpy/pandas per the competition rules.

Python numpy pandas

The code for this work belongs to the companies, so it isn't public. Happy to walk through the design.

• Real-time data platform: unified Polymarket and Kalshi feeds into one ML-ready stream at 40M+ messages/hour (Kafka, WebSockets, Protobuf), with MCP servers giving AI agents live market data as tools
• LLM-as-judge evaluation: a GPT-4 judging framework across 5+ question types that raised agreement with human reviewers from 60% to 90%
• Autonomous trading agent: real-time data → ML models → execution in under 1 second, with risk limits and a kill switch in deterministic code

Have an AI project in mind? See case studies at bprakashputta.github.io, or message me on LinkedIn or Upwork with what you're trying to build, and I'll suggest the simplest approach that works.