Ask the portfolio

VC
VC
chanamalluvinay@gmail.com
The 30-second version

Vinay Chanamallu

Below the surface, the sub's cutaway shows Vinay giving orders on the bridge, his robot at the sonar, and four robot crew members wearing the OpenAI, Anthropic, Google DeepMind and xAI logos.

A chrome submarine at the surface. Vinay, in a white captain’s cap, looks out from its hatch, shading his eyes.

Generative AI / LLM Engineer

Anyone can get a model to produce an answer.The engineering is in knowing whether that answer is right.

Retrieval in production at GOML

01Retrieve

A corpus is just noise

Sixty-five thousand documents. Every one of them equally likely to be the answer, which is another way of saying none of them are.

65K documents, 850+ queries a day

02Score

Score every candidate

BM25 and vector search put a number on relevance. The field stops being flat. Most of it is still wrong, and that is the point.

BM25 + vectors

03Re-rank

Rank, then re-rank

A cross-encoder reorders the shortlist. This is where the retrieval metric actually moves, and where most systems stop too early.

Recall@5 72% to 84%

04Ground

Keep five. Cite them.

The generator only sees what survived. Hallucinated answers fall by a fifth because there is less left to hallucinate from.

hallucinations -20%

A diagram of the retrieval pipeline behind GOML’s multi-tenant GenAI platform: a corpus of document chunks is scored for relevance, ranked, re-ranked by a cross-encoder, and cut to the top five results that are passed to the generator.

Three systems, and the decision behind each

01Production ML

4 of 5churners caught

A chrome birdcage with its door open. Chrome fish swim out of it; orange threads catch four of the five and pull them back.

Illustration. Real model: 0.853 recall, 0.860 ROC AUC.

Telco Churn

A retention model tuned on money, not on accuracy.

0.860
ROC AUC
0.853
Recall
1.28x
Return
Problem
A telecom operator was losing subscribers with no way to see it coming. Retention budget went out uniformly, which means most of it reached people who were never going to leave.
System
A gradient-boosted classifier over behavioural and billing features, served from a BigQuery feature store on Cloud Run. SHAP values ship with every prediction, so the marketing team sees why an account was flagged and not just that it was.
Decision
Accuracy was the wrong objective. A missed churner costs far more than a wasted retention offer, so the decision threshold is tuned on expected value rather than F1. Drag the cutoff in the panel to see the trade: push it left and you catch more churners while burning budget on people who would have stayed.
Result
0.860 ROC AUC at 0.853 recall, and 1.28x return on retention spend against the uniform baseline. A live dashboard ranks at-risk accounts daily.
0%
Churners caught
0
Accounts flagged
0%
Spend wasted

Move the cursor across the field to set the cutoff

Panel illustrates the threshold mechanism on a synthetic cohort. The metrics beside it are measured on the real model.

  • XGBoost
  • SHAP
  • Cloud Run
  • BigQuery
  • Vertex AI
GitHub ↗Live dashboard ↗

02Agentic systems

Five chrome robots and a chrome octopus dealer sit round a blue poker table. One robot hides a glowing orange card under the table.

Illustration of the protocol: every move is JSON against a schema, streamed over SSE.

Poker Agents

Five local models at one table, with an incentive to lie.

5
Agents
0
Cloud calls
SSE
Transport
Problem
Most multi-agent work is evaluated on cooperative toy tasks. I wanted a setting with hidden information, real stakes between agents, and a reason to misrepresent, to see how small local models actually behave.
System
Five Ollama-backed agents share a table through a FastAPI game engine. Every action is emitted as structured JSON against a schema, hands are scored with Treys, and the match streams to the browser over server-sent events.
Decision
Agent reasoning is constrained to a typed schema instead of free text. It costs some expressiveness and buys a parser that never fails mid-hand. That is the same trade I would make for any tool-calling agent going to production.
Result
Runs entirely on local models with no API cost. The turn-taking, message passing and failure modes map directly onto enterprise multi-agent workflows.

Real capture, not a mockup

Recorded locally, no cloud inference

  • Python
  • FastAPI
  • Ollama
  • SSE
  • Treys
GitHub ↗Write-up ↗

03Published research

A city skyline at dusk across still water, its towers rising to a peak right of centre. As the forecast line crosses the rooftops, the lights come on.

The skyline is the forecast: an ensemble of deliberately uneven models.

Grid Demand

An ensemble that wins by being deliberately uneven.

-16%
RMSE
-20%
MAE
IEEE
Published
Problem
Grid operators schedule generation against a demand forecast. Single-model forecasts plateau, and the leftover error turns into either idle capacity or a shortfall.
System
An ensemble over gradient-boosted learners, with feature engineering for daily and weekly structure plus weather coupling. Each member is biased in a different direction so that their errors partly cancel.
Decision
The members are not individually optimal, on purpose. Diversity across the ensemble beats per-model accuracy, which is why the combined residual collapses further than either member reaches alone.
Result
16% lower RMSE and 20% lower MAE against the baseline. First-authored and accepted at IEEE CICT 2025.

RMSE 0.0610/errors cancel, so the pair beats either alone

Panel illustrates error cancellation on a synthetic series. The RMSE and MAE figures are the published results.

  • XGBoost
  • LightGBM
  • Forecasting
  • Research
Read the paper ↗

Everything else

Research pipelines, competition models and side builds. Shorter stories, same habit of measuring first.

  1. 01Crash Clearance Modelling

    Transportation safety research with the CATT Lab at UMD. Pipeline reduces collinear predictors by variance inflation before fitting.

    6x R² gainXGBoostOpen project ↗
  2. 02Pneumonia Detection

    Chest X-ray classification on transfer-learned EfficientNetB0, with augmentation tuned for the class imbalance in the source set.

    95.2% AccuracyTensorFlowOpen project ↗
  3. 03Airbnb Booking Rates

    Competition-winning stacked ensemble over listing, host and review features, with SHAP attributions for the top predictors.

    0.912 ROC AUCXGBoost, CatBoostOpen project ↗
  4. 04Flight Delay Pipeline

    End-to-end pipeline over Bureau of Transportation Statistics records, joined against historical weather at origin and destination.

    3.6M RecordsXGBoost, CLIOpen project ↗
  5. 05Walmart Inbound Forecasting

    UMD capstone with Walmart. Twenty-eight day case and truck volume forecasts driving dock and labour planning.

    30% Over baselineXGBoost, StreamlitOpen project ↗
  6. 06RAG Chatbot

    Retrieval-augmented question answering over a document corpus, with chunking and reranking isolated so each can be evaluated alone.

    Python, LLMOpen project ↗
All repositories ↗Tableau Public ↗

Before the models, the analytics

Twelve published Tableau workbooks. Most of what I know about framing a question for a model started here.

US housing market, Redfin
US housing market, RedfinTableau
Airbnb supply across New York
Airbnb supply across New YorkTableau
Streaming catalogue overlap
Streaming catalogue overlapTableau
US domestic air traffic
US domestic air trafficTableau
Japan growth and innovation
Japan growth and innovationTableau
Goodreads ratings behaviour
Goodreads ratings behaviourTableau
Global video game sales
Global video game salesTableau
Wage distribution analysis
Wage distribution analysisTableau

Experience

Research to production, three years and counting.

2026

May 2026 to Present

AI Engineer

GOML

Owns retrieval, LangGraph orchestration and evaluation for a multi-tenant GenAI platform on AWS Bedrock: 3 client teams, 850+ queries a day across 65K documents. Hybrid search (BM25 + Pinecone, Cohere reranking) measured by an LLM-as-judge evaluation pipeline, a LangGraph onboarding workflow that cut analyst handling time 30%, tenant-scoped Redis caching that cut cost per query 20% and p95 latency 16%, and tenant isolation enforced outside the LLM.

  • LangGraph
  • AWS Bedrock
  • Pinecone
  • Redis
  • ECS Fargate
  • Terraform
84%
Recall@5, from 72%
-20%
Hallucinated answers

2025

Sep 2025 to Feb 2026

AI Engineer Intern

GOML

Document Q&A proof of concept (RAG) for a professional-services client across 4,800 contracts, SOWs and process manuals. Built the ingestion and retrieval pipeline and tuned chunking, retrieval and prompts; earned a return offer after the proof of concept went to production.

  • RAG
  • Document ingestion
  • Retrieval
  • Prompt engineering
3.9/5
Answer quality, from 3.4
4,800
Documents

2023

Aug 2023 to Aug 2024

AI/ML Engineer

Brillio

Production NLP ticket routing for a financial-services client: fine-tuned DistilBERT into 12 categories, served through FastAPI on AWS ECS at 5,000+ requests a day, with automated training, evaluation and deployment that shortened releases from 10 to 7 business days. Also prototyped BART summarization and an early RAG Q&A tool that cut median analyst review time 25%.

  • DistilBERT
  • FastAPI
  • ONNX Runtime
  • SageMaker
  • MLflow
0.80
Macro F1, from 0.73
-43%
p95 latency

2021

Oct 2021 to Jul 2023

Data Scientist

Wipro

Store-SKU demand forecasting for a retail client with ~100 stores: XGBoost models forecasting weekly demand up to four weeks ahead from 8.4M sales line items. An 80-store randomized pilot of forecast-driven replenishment showed a 13% lower stockout rate, and PySpark and Airflow jobs feeding Tableau and Power BI cut weekly manual reporting by more than half.

  • XGBoost
  • PySpark
  • Airflow
  • Tableau
  • Power BI
-10%
WAPE
-13%
Stockout rate

Evaluation is the whole job.

New York open to relocation

Vinay holds up a scorecard and gives a thumbs up; the robot cheers.

Anyone can get a model to produce an answer. The engineering is in knowing whether that answer is right, what it costs when it is wrong, and how you would notice if it quietly stopped working.

I work across retrieval, agents, NLP and forecasting, and the pattern repeats every time. Build the measurement first. The model is the easy part.

Based in New York. Currently on retrieval, LangGraph orchestration and evaluation for a multi-tenant GenAI platform on AWS Bedrock. Open to relocation or remote.

GenAI and agents

  • LangGraph
  • RAG
  • AWS Bedrock
  • Pinecone
  • LLM-as-judge evaluation
  • Guardrails
  • Tool calling
  • MCP
  • Ollama

Platform and data

  • AWS ECS
  • SageMaker
  • Terraform
  • Vertex AI
  • Cloud Run
  • BigQuery
  • Docker
  • Airflow
  • FastAPI
  • PySpark

Modelling

  • Python
  • SQL
  • XGBoost
  • LightGBM
  • PyTorch
  • TensorFlow
  • Hugging Face
  • SHAP
  • MLflow
  • A/B testing

Open to GenAI, LLM and ML engineering roles

Let’s build something measurable.

chanamalluvinay@gmail.comEmail me