Skip to content
Nikhil Kumar Reddy

AI/ML Engineer · Washington, DC

I build ML systems that survivecontact with production.

Three years taking LLM and ML systems from prototype to production across healthcare and enterprise — RAG pipelines, fine-tuned models, multi-agent applications, and voice agents where a wrong answer has consequences.

Open to ML Engineer, AI Engineer and Applied Scientist roles

Every row links to the work behind it.

01Selected work

Five systems, and what broke in each.

Every case study names a real limitation. A wall of green checkmarks tells you nothing about whether someone has run a system in production.

guardrail2026

Post-Discharge Voice Agent

Calls patients 72 hours after discharge — and escalates on hardcoded clinical rules the AI is never allowed to touch.

100%
Sensitivity across 92 evaluation cases
0%
False-escalation rate
418
Passing tests
~0.7s
In-call LLM response

  1. Twilio outbound

    Human-in-the-loop

    ~72h post-discharge

  2. Deepgram STT

    Retrieval

    streaming transcription

  3. Layer A — raw scan

    Guardrail

    pre-LLM, on fragments

  4. Claude Haiku

    Generation~0.7s

    fills slots only

+ 4 more stages

Read the case study →

human2026

Inbound Clinic Voice Agent

A real phone number that answers, books an appointment end to end, and hangs up.

Live. Call it and book an appointment.

88% → 100%
Task completion
19/19
Evaluation cases passing
P50 4.1s · P95 5.2s
End-to-end turn latency
P50 87ms · P95 171ms
ASR latency

  1. LiveKit SIP

    Human-in-the-loop

    inbound PSTN

  2. Deepgram ASR

    RetrievalP50 87ms

  3. Claude Haiku 4.5

    GenerationP50 2,925ms

  4. FastAPI tool calls

    Human-in-the-loop

    real availability + booking

+ 2 more stages

Read the case study →

retrieval2026

SEC Finance RAG

Citation-backed answers over 25,000 SEC filings, with the guardrails that make them trustworthy.

<500ms
Median query latency, cache miss
>0.80
RAGAS faithfulness
<0.20
DeepEval hallucination rate
60–70%
Production cache hit rate

  1. Async EDGAR ingest

    Retrieval

    aiohttp, concurrent

  2. Format-aware parse

    Retrieval

    iXBRL · PDF

  3. Sentence-window chunk

    Retrieval

    tables kept atomic

  4. Hybrid index

    Retrieval

    BM25 + BGE-large 1024d

+ 7 more stages

Read the case study →

retrieval2026

Vectorless RAG

Retrieval over SEC filings with no embeddings and no vector database at all.

top-K → 1–3 whole
Sections retrieved per query
9–12s
Response time
6 (FY2020–25)
Filings indexed

  1. Tree index

    Retrieval

    HTML → Markdown → JSON tree

  2. Node summaries

    Generation

    title + summary per node

  3. Structural reasoning

    Generation

    Claude reads the tree

  4. Targeted fetch

    Retrieval

    1–3 whole sections

+ 1 more stages

Read the case study →

generation2026

AnalystAI

Seven specialised agents that take a spreadsheet and a plain-English question to a boardroom answer.

7
Specialised agents in the pipeline
E2B sandbox
Generated code isolation
CSV · Excel · SQL · Sheets · REST
Supported sources

  1. Intake

    Generation

    question → analysis plan

  2. Ingestion

    Generation

    load + infer schema

  3. Cleaning

    Generation

    nulls, dupes, coercion

  4. E2B sandbox

    Guardrail

    generated Python, isolated

+ 4 more stages

Read the case study →

All 14 case studies →
02Capability

Skills, arranged as the pipeline they belong to.

A badge wall tells you what someone has touched. This shows where each tool sits in a system that has to work on Monday.

Ingestion & retrieval

Getting the right context in front of the model — which is where most RAG systems are won or lost.

  • LangChain
  • LlamaIndex
  • FAISS
  • Pinecone
  • ChromaDB
  • Qdrant
  • BM25 hybrid search
  • BGE / MiniLM embeddings
  • Cohere Rerank
  • PySpark
  • Databricks
  • Snowflake
  • dbt

Modelling & generation

Fine-tuning where it earns its cost, prompting where it does not, and orchestration when one call is not enough.

  • PyTorch
  • TensorFlow
  • Hugging Face Transformers
  • BERT
  • Llama
  • LoRA / QLoRA / PEFT
  • Claude API
  • OpenAI API
  • LangGraph
  • CrewAI
  • AutoGen
  • MCP
  • XGBoost
  • scikit-learn

Safety & guardrails

The rules that live outside the model, because anything a model can generate it can eventually generate wrongly.

  • Input validation
  • Prompt-injection filtering
  • PII redaction
  • Output verification
  • Hallucination checks
  • Sandboxed execution (E2B)
  • Hardcoded escalation rules

Evaluation & observability

If it is not measured it is not engineering. Every system above ships with the harness that proves it works.

  • RAGAS
  • DeepEval
  • LangSmith
  • MLflow
  • A/B testing
  • Drift detection
  • SHAP
  • Held-out evaluation sets
  • Model monitoring

Deployment & the human loop

Shipping it, keeping it up, and deciding deliberately where a person stays in the path.

  • Docker
  • Kubernetes
  • FastAPI
  • AWS SageMaker
  • Azure ML
  • GCP Vertex AI
  • Airflow
  • CI/CD
  • Railway
  • PostgreSQL
  • MongoDB
03Experience

Three roles, each one a deliberate step.

Data analysis to NLP modelling to machine learning engineering. The numbers here are employer-attested — the ones tied to public code live on the case studies instead.

AI/ML Engineer

Milken Institute School of Public Health, GWU

Owned the full ML lifecycle in a domain where model reliability is not optional: retrieval pipelines, fine-tuning, inference endpoints, and production systems with automated retraining.

  • Built a LangChain RAG pipeline over 15K behavioural health records using GPT-4, measuring retrieval quality with RAGAS rather than eyeballing outputs — a 34% improvement over the keyword baseline, which let clinical teams handle 3x more patient queries.
  • Fine-tuned BERT via LoRA on 5K patient survey responses for clinical text classification, taking held-out accuracy from 71% to 87%, then sat with clinical staff to validate outputs and document the edge cases the metric hid.
  • Containerised FastAPI inference endpoints on AWS SageMaker with Docker, adding input validation and output safety constraints before anything reached a decision-support workflow. Average response latency fell 45%.
  • Automated retraining with MLflow, drift detection and CI/CD guardrails, replacing an ad-hoc process and sustaining 78% production uptime across model versions.

Jan 2025 – May 2026Washington, DC

71% → 87%
Clinical text classification accuracy
baseline → +34%
RAGAS retrieval quality vs keyword baseline
baseline → −45%
Inference response latency

Data Scientist (NLP/ML)

Data Science for Sustainable Development

The transition point from traditional data science into NLP-heavy, production-oriented work — semantic search and automated feature engineering for policy research teams.

  • Deployed FAISS semantic search with anomaly validators across JSON, CSV and Parquet census data, tracing cross-format quality gaps with the data engineering team until the pipeline reached full compliance.
  • Designed automated feature engineering workflows in PySpark on Databricks, using LangGraph agents to surface and test candidate features — 35% more modelling throughput for a policy research team.
  • Shipped Streamlit dashboards backed by Azure OpenAI that translated model output into plain-language summaries for non-technical leadership, cutting reporting prep by 40%.

Aug 2024 – Dec 2024Washington, DC

baseline → +35%
Statistical modelling throughput
baseline → −40%
Stakeholder reporting prep time

Data Scientist

Cogno AI

Two years across e-commerce, starting in analytics and progressively taking on model building, deployment and MLOps until I owned pipelines end to end.

  • Built a production RAG system on OpenAI embeddings and Pinecone across 10K e-commerce SKUs, taking average query latency from 3.2s to 0.8s — the difference between search that felt broken and search that felt instant.
  • Trained an XGBoost demand forecasting model on retail POS data, 17% more accurate than the ARIMA baseline, reducing annual inventory overstock.
  • Set up the team's MLOps foundation — MLflow, Git, Docker, experiment tracking, A/B model evaluation and observability dashboards — cutting the iteration cycle by 45% and becoming the standard the team kept.

May 2022 – Jul 2024India

3.2s → 0.8s
RAG query latency
baseline → +17%
Demand forecast accuracy vs ARIMA
baseline → −45%
Model iteration cycle

MS, Data Analytics · The George Washington University · Aug 2024 – May 2026

BTech, Computer Science & Engineering · Dayananda Sagar University

04About

How I got here, and how I work.

I spent three years moving from data analysis to NLP modelling to machine learning engineering, and the move was deliberate. I wanted to understand not just how models learn, but how they hold up once real people depend on them — with messy inputs, latency budgets, and users who do not care about your F1 score.

Most of what I know now came from the gap between those two things. A retrieval pipeline that scores well offline and falls apart on a follow-up question. A classifier that is accurate until a clinician asks why. A voice agent where the interesting engineering is not the model at all, but the rules you keep outside it so a bad generation can never reach a patient.

So my work tends to look like this: build the thing, instrument it, find where it breaks, and put a guardrail there. Every project below shows that trace — including the parts that did not work.

05Contact

Currently looking for ML and AI engineering roles.

If you are hiring, or you just want to argue about whether vectorless retrieval is a good idea, I would like to hear from you.