Open to Full-Time Roles

Naga Prem Sai
Pendela

AI Engineer & Applied AI Engineer building production-grade multi-agent systems and decision-intelligence platforms. MEng CS @ University of Cincinnati (GPA 3.96).

4x
Agents in NexusIQ — SQL, RAG, Web, Fusion
534K
Transactions analyzed in RevenueIQ
97.7%
RAG Hit@5 retrieval accuracy
95%
Churn prediction F1-score

Work

Featured Projects

Live · AWS EC2 · Docker · CI/CD · May 2026
NexusIQ AI

A 4-agent business intelligence platform that routes questions across SQL, documents, and live web data — then cross-validates answers from different sources and tells you how confident it is.

97.7%
RAG Hit@5 retrieval accuracy
0.03%
SQL ↔ PDF cross-validation delta
5–12s
Multi-source validated query time
~60%
Latency cut via parallel agents

Before
An analyst answering "Why did West region revenue drop?" manually queries the database (45 min), reads PDF reports (2 hrs), checks competitor sites (1.5 hrs), and writes a summary. Total: 2–3 days of work.
After
NexusIQ routes the question to all 3 sources simultaneously, cross-validates that database and PDF numbers agree within 1%, and returns a cited, confidence-scored answer in 5–12 seconds.

Query router
SQL agent
+
RAG agent
+
Web agent
Fusion agent
Cited answer

Agents run in parallel via ThreadPoolExecutor. Fusion cross-validates SQL numbers against PDF text — within 1% = HIGH confidence, deterministic formatting (no LLM call). Results traced to AWS CloudWatch + local JSONL. 100,000-row Supabase DB · 43 PDFs (425 ChromaDB chunks) · 5 live product categories via 4 scraping techniques.


Vector search ranked the wrong document #1
❌ Failed

A "Q3 revenue" query returned a "Q4 Outlook" chunk ranked first — the Q4 section mentioned revenue more often, so cosine similarity put it top. Tried ChromaDB metadata filtering ($contains) — not supported. Tried filename-based filtering — missed data inside differently-named files.

✓ What worked

Hybrid BM25 + vector search. BM25 catches exact tokens ("Q3", "2024"), vector catches meaning. Accuracy jumped 33% → 100% on the benchmark set. The hybrid_search() function existed in the codebase but was never called — wiring it in was the fix.

Parallel agents: asyncio vs. multiprocessing vs. ThreadPoolExecutor
❌ asyncio

All agent methods are synchronous. Full async rewrite = 500+ lines changed and introduced a Streamlit event loop conflict.

❌ multiprocessing

Can't pickle LangChain objects. Each process re-initialises SentenceTransformer (80MB, 5–10s per process).

✓ ThreadPoolExecutor

All agents are I/O-bound. GIL doesn't block I/O threads. Zero agent code changes. ~60% latency reduction on multi-source queries.

HIGH confidence answers — skip the LLM entirely
✓ Deterministic formatting

When SQL and PDF numbers match within 1%, the Fusion agent formats the answer directly — 0 LLM calls, ~0.3ms. LLM synthesis only fires for conflicts or low-confidence cases. This made answers faster, more stable, and ~40% cheaper in token cost. The insight: the LLM is for resolving uncertainty, not for presenting facts that are already validated.


Retrieval accuracy
97.7% Hit@5
0.919 Context Recall. Cross-encoder reranker on top of hybrid BM25 + vector search. Benchmarked on 43-query golden eval set.
Cross-validation precision
0.03% delta
SQL $15,399,999 vs PDF "$15.4M" — validated HIGH confidence. Eliminates manual fact-checking across sources.
Test coverage
194 unit tests
+ 12 golden eval cases, 7 offline evals, 43-query RAG benchmark. Tested like a production service, not a demo.
Fault tolerance
Circuit breaker
Gemini → Groq automatic fallback. Zero query failures during recruiter testing. Degraded mode discloses which source was unavailable.

Add a FastAPI layer — current Streamlit UI is demo-grade. A REST API makes NexusIQ embeddable in enterprise tools without the Streamlit dependency.
Compute cost-per-query from the LLM gateway ledger — token data exists, needs a dollar conversion. "$0.003 per multi-source BI query" is a CFO-readable metric.
OCR support for scanned PDFs — current pipeline handles text-native PDFs only. Adds an entire class of real enterprise documents.
Migrate ChromaDB into pgvector on Supabase — eliminates a second infrastructure dependency and simplifies the production stack.

LangGraph ChromaDB Hybrid BM25 + vector Cross-encoder reranker Gemini 2.5 Flash Groq LLaMA 3.3-70B PostgreSQL · Supabase Python · Pandas AWS EC2 · ECR · S3 Docker GitHub Actions CI/CD Caddy · HTTPS CloudWatch LLM gateway telemetry JSONL traces
Live · Render · Streamlit Cloud · March 2026
RevenueIQ AI

An ML-powered decision intelligence platform that transforms 534K e-commerce transactions into actionable retention, growth, and forecasting insights — in 30 seconds instead of 4 hours.

95%
Churn prediction F1-score
$4M+
Revenue opportunities surfaced
30s
Full executive report time
7.74×
Query speedup over Pandas

Before
E-commerce teams waste 4–6 hours every week manually building revenue reports in spreadsheets. By the time the report is written, the opportunities it identifies have already been missed.
After
RevenueIQ runs 5 ML models on 534K transactions, identifies 978 at-risk customers, forecasts next 30 days revenue, and generates a plain-English executive summary — in 30 seconds.

534K transactions
DataCleaner
5 ML models
DuckDB
Groq LLM
Streamlit

5 models: Random Forest churn (95% F1) · KMeans segmentation (k=5) · Isolation Forest anomalies · Prophet + ARIMA/ETS forecasting · RFM scoring. ML results stored back to DuckDB as queryable SQL tables — the entire team can query segments without touching Python. Groq LLM (LLaMA 3.3-70B) auto-generates 5 executive insight types from model outputs.


Churn model showed 100% accuracy — which was wrong
❌ First version

Initial Random Forest hit 100% accuracy on the test set. Immediately suspicious. Investigation found data leakage: recency features were computed using the full dataset before the train/test split, so the model saw future information during training.

✓ Fixed version

Recomputed recency features strictly from training-period data. Accuracy dropped to 95% F1 — still strong, now trustworthy. The lesson: a model that's too good is wrong, not exceptional. 95% on a clean split is worth more than 100% on a leaky one.

KMeans over RFM — and why k=5 not k=7

Traditional RFM requires manually setting thresholds and uses only 3 features. KMeans discovers natural clusters using 6 features: recency, frequency, monetary value, AOV, lifespan, quantity. This separated 18 ultra-VIP customers worth $96,614 each — customers RFM would have grouped with ordinary Champions.

k=5 rationale

k=7 had a 5.5% better silhouette score. But k=2 → k=5 gave a 50% improvement. Diminishing returns. Five segments map cleanly to actionable personas: VIPs, Loyal, Potential Growers, At-Risk, Lost. Seven creates splits between similar groups that no marketing team will act on distinctly.

DuckDB over Pandas — and why it mattered
❌ Pandas on repeat dashboard loads

Every filter or refresh re-read and re-aggregated the full 534K row CSV. Dashboard cold load: 45 seconds. No columnar storage, no query cache, no optimisation.

✓ DuckDB + st.cache_data

DuckDB is OLAP-optimised — columnar storage, vectorised execution, embedded (zero server). Combined with @st.cache_data, repeat loads dropped from 45s to ~3s. 7.74× faster queries. ML results stored in DuckDB so the whole team can run SQL against segment data without Python.


Customer segmentation
5 KMeans personas
18 VIP Champions ($1.74M revenue, $96K each avg). 1,692 Potential Growers. 978 At-Risk flagged for win-back.
Revenue forecasting
$1.61M / 30 days
Prophet model, 9% MAPE (91% accuracy). Exponential Smoothing MAE $18,365. CFO-ready 4-week revenue outlook.
Anomaly detection
5,249 flagged
Isolation Forest. Anomalies averaged 230 units vs. normal 8 — identified as B2B/wholesale. Wholesale pricing opportunity surfaced.
Report generation
2 hrs → 30 sec
Groq LLM auto-generates executive summary, customer personas, forecast commentary, anomaly explanation, and churn playbook.

Connect to a live Shopify or WooCommerce API — current pipeline reads from a static CSV. One file-path change connects it to live data.
Add cohort-based revenue forecasting — current Prophet model forecasts total revenue. Per-segment forecasting shows which cohort is driving growth vs. decline.
Build a campaign ROI tracker — feed win-back campaign results back into the churn model to measure actual recovery vs. predicted.
Replace Streamlit with FastAPI + React frontend — makes the platform embeddable in existing business tool stacks.

Random Forest KMeans Isolation Forest Prophet ARIMA / ETS Scikit-learn Python · Pandas · NumPy DuckDB Groq LLaMA 3.3-70B Streamlit Plotly Render

Background

Experience

Research Data Analyst & Team Lead
Center of Innovation & Entrepreneurship, VCE · Hyderabad, India
Aug 2023 – Apr 2024

Education

Education

Master of Engineering, Computer Science
University of Cincinnati · Cincinnati, OH
Aug 2024 – May 2026GPA 3.96
Focus: Data Analytics, AI Systems, Machine Learning
B.Tech, Computer Science & Engineering
Vardhaman College of Engineering · Hyderabad, India
Dec 2020 – May 2024GPA 3.78
IEEE-published research in NLP · Cyberbullying detection

Learning

Certifications

📊
Quantium Data Analytics Job Simulation View Certificate ↗
Forage · March 2026
  • Completed an industry simulation focused on data analytics and commercial insights for a data science team
  • Extracted customer and transaction insights to deliver data-driven commercial recommendations
  • Identified benchmark stores for uplift testing on trial store layouts using statistical analysis
  • Built comprehensive reports for a Category Manager enabling evidence-based strategic decisions
Data Analysis Data Cleaning Statistical Analysis Data Visualization Commercial Thinking

Toolkit

Skills

AI / ML
Multi-Agent AI RAG LLM Orchestration Hybrid Search LangGraph NL→SQL Machine Learning Prompt Engineering
Data & Analytics
Python SQL DuckDB Pandas NumPy Scikit-learn Plotly Streamlit
Databases
PostgreSQL ChromaDB Supabase Vector DBs SQLAlchemy DuckDB
Cloud & DevOps
AWS EC2 AWS ECR AWS S3 CloudWatch Docker GitHub Actions CI/CD Caddy
LLM APIs & Tools
Groq Gemini LLaMA HuggingFace BeautifulSoup Selenium

Let's Connect

Open to AI/ML Engineer, Applied AI Engineer, and Forward Deployed Engineer roles