name:     Nihanth Naidu Kalisetti
role:     AI Engineer
based:    Brooklyn, New York
studying: MS Artificial Intelligence
          LIU Brooklyn (Dec 2026)

builds:
  - multi-agent orchestration
  - production RAG pipelines
  - LLM evaluation + tracing
  - MLOps + inference infra

principle: trust over capability
shipped:   2 PyPI packages
writing:   LinkedIn Top AI Voice
status:    open to AI roles

Open source. Each ships with a fixed benchmark and documents its own failure modes on purpose, because hiding them would defeat the point.

LangGraph · PostgreSQL · pgvector · FastAPI · Next.js · Claude · OpenAI

A read-only, database-agnostic natural-language to SQL agent. It is read-only by construction, not by convention: the database is reached only through a PostgreSQL role granted nothing but SELECT, so a write is rejected twice, once by the API guard and definitively by the role itself. There is no write mode. A 14-node graph classifies scope and refuses out-of-scope, write, and irreducibly ambiguous requests up front, retrieves schema and few-shot examples through pgvector, verifies that generated SQL references only tables and columns that exist before execution, and self-corrects failed queries under a bounded retry.

Two failures are left unfixed and documented: a faithfulness gap where a well-formed grounded query can quietly narrow the request, and a DISTINCT omission on some queries. Both are architectural boundaries, stated plainly rather than papered over.

LangGraph · pgvector · BM25 · Redis · FastAPI · Claude Haiku · Tavily

A 13-node cyclic pipeline with a self-correcting hallucination loop. Every answer is scored on faithfulness, relevancy, and groundedness. If faithfulness drops below 0.75, the query is rewritten and retrieval runs again, up to three times. The pipeline routes across three retrieval strategies (BM25, pgvector, and RRF hybrid fusion) by query type, reranks candidates with a cross-encoder before generation, and falls back to live web search when the corpus returns nothing, rather than burning correction cycles on a futile loop. A two-tier Redis semantic cache serves sub-3-second responses on hits versus 30 to 90 seconds for a full run.

Python · Docker · radon · vulture · ast · Claude Opus

Darwin takes one working-but-bloated Python unit, has a language model simplify it, then objectively verifies the result is correct, behavior-preserving, and structurally simpler, or returns the original unchanged. The point of the project is the verification, not the rewrite. No language model judges correctness or quality anywhere in the scoring path. A candidate must pass a correctness gate (pytest inside a locked-down container), a behavior gate (exact reproduction of withheld golden input-output pairs, or differential equivalence at the public API), and only then receives a structural complexity score computed by radon, vulture, and ast.

Two guarantees define it. It is never worse than baseline: the loop returns the best gate-passing candidate or the unmodified original. And it is access-isolated: every candidate runs with no network, no host filesystem, and capped memory, CPU, and PID as a non-root user. One residual is documented honestly rather than hidden: in-process verdict forgery stays open by deliberate decision, deferred to a future process split. Darwin is an access-safe verifier, not a forgery-proof judge, and it does not claim otherwise. Validated on a 30-case single-function corpus plus 3 multi-file cases, 150 runs, 95.3% reaching the returned-best candidate in generation one.

LangGraph · GPT-4o · Tavily · FastAPI · React 19 · PostgreSQL · ReportLab

Six specialized agents orchestrated by a deterministic supervisor. No language model is involved in any routing decision; the supervisor routes purely on explicit state-inspection rules.

• Parallel fact-checking through the LangGraph Send() API fans out all claim verifications simultaneously, bringing verification time to roughly three seconds regardless of claim count.
• Human-in-the-loop outline approval via interrupt() with PostgresSaver checkpointing pauses the graph mid-run and resumes from the exact node without re-running completed phases.
• Report versioning diffs the original outline against the approved one and skips synthesis for unchanged sections, saving around 80% of tokens when only part of a report changes.

pip install inputguard

pip install ai-stamp

SpectraVoice — a hands-free, screen-aware voice assistant for macOS that sees the screen in real time, controls keyboard and mouse, searches the web, and responds with natural text-to-speech. Supports both cloud and local Ollama inference with barge-in interruption. Python · Whisper · Ollama · FastAPI

_{More projects, including live demos, are pinned below and on the portfolio.}

AI · GenAI · Agents

MLOps · Infrastructure

Data · Vector Search · ML

Languages

AI Engineer, Intern  ·  Cogent LLC, Remote (Nebraska)  ·  Jun 2026 – Present Building an AI data platform that unifies fragmented structured and unstructured sources into one queryable layer, with LLM-powered natural-language access over enterprise data. Designing hybrid RAG retrieval (pgvector with keyword search and reranking) behind FastAPI services so teams query data in plain language instead of hand-written SQL, and building ingestion and embedding pipelines that turn raw records and documents into versioned, searchable knowledge with evaluation and guardrails that keep responses grounded and auditable.

Founding AI/ML Engineer, Contract  ·  UnityGrid AI  ·  Jul 2024 – Jan 2025 Designed multi-agent LangChain workflows for semiconductor design validation, cutting manual review time by 36% across 3 engineering teams. Engineered a pgvector and Redis RAG system that brought p95 latency under 900ms and lowered infrastructure cost by 28%. Established LangSmith observability with continuous hallucination tracking, shifting regression detection from post-deployment to pre-release.

AI/ML Engineer, promoted from ML Engineer  ·  Futzen eTechnologies  ·  Apr 2021 – Jun 2024 Promoted to AI Engineer within 2 years. Built customer segmentation and behavior models with K-Means, decision trees, and collaborative filtering, deployed across live production pipelines. Streamlined feature-engineering pipelines, cutting model training prep time by 35% and lifting F1 across production deployments.