ROHITH KUMAR REDDIPOGULA
OPEN TO WORK — BERLIN
> retrieve candidate WHERE role = "AI/NLP Engineer" AND location = "Berlin"

I build systems that
find the right answer
among millions.

MSc Data Science, University of Europe for Applied Sciences, Potsdam. My thesis discovered the optimal fusion weight for hybrid retrieval — alpha = 0.70, not the commonly assumed 50/50 split — validated against 8.84M passages. Since then I've shipped eight production AI systems, alone, end to end.

5 systems live right now Job Seeker Visa EU Blue Card eligible Available immediately
Recall@10
93%
on 8.84M MS MARCO passages
Significance
p=0.002
paired t-test, alpha=0.70
Fine-tuning
0.089%
params trained, loss 2.47→0.89
Faithfulness
0.909
RAGAS, Gemini-as-judge
// ABOUT

I research how retrieval actually works, then ship it. Most hybrid RAG tutorials assume sparse and dense retrieval should be weighted 50/50. My thesis tested that assumption directly — eleven tracked experiments later, the real optimum was alpha=0.70, validated with a paired t-test (p=0.002). That single number improved Recall@10 by 11.4% over the naive baseline.

After graduating I kept building instead of stopping at the thesis: fine-tuning LLMs with QLoRA on a free Colab GPU, orchestrating four-agent LangGraph pipelines, deploying to AWS EC2 with systemd, running Kubernetes with zero-downtime rollouts, and tracking every experiment publicly on MLflow.

I'm now looking for an ML Engineer, NLP Engineer, or AI Engineer role in Berlin. Job Seeker Visa holder, EU Blue Card eligible, available immediately.

Core ML & NLP
PythonPyTorchHuggingFaceRAGFAISSBM25E5-base-v2
LLM stack
LangChainLangGraphQLoRARAGASGeminiGroqUnsloth
Infrastructure
DockerKubernetesAWS EC2FastAPIsystemdBand SDK
MLOps
MLflowDagsHubHF SpacesGit

Live systems

8 results, sorted by relevance
// query: "production-grade AI systems built solo" — 0.04s
01thesis
BM25 sparse search fused with Microsoft E5 dense embeddings across 8.84M MS MARCO passages. The fusion weight (alpha=0.70) was discovered through eleven tracked experiments, not assumed.
93% Recall@10 · +11.4% vs baseline · MRR=1.0
FAISSBM25E5FastAPIDocker
0.98
relevance
02agent
Four specialised LangGraph agents — Search (DuckDuckGo), Summarise (temp 0.3), Fact-Check (temp 0.1), Writer (temp 0.4) — each tuned for one job in the pipeline.
4 agents · stateful LangGraph · 3 distinct temperatures
LangGraphGeminiDuckDuckGoDocker
0.95
relevance
03agent
A LangGraph ReAct agent holding three tools — web search, calculator, RAG retrieval — and reasoning about which one a given question actually needs.
3 tools · autonomous tool selection
LangGraphGeminiStreamlit
0.93
relevance
04tuning
4-bit quantisation with LoRA adapters on a free Colab T4 GPU. Trained on NLP domain knowledge, published openly on HuggingFace Hub.
0.089% of params trained · loss 2.47 → 0.89
QLoRAUnslothTinyLlama
0.91
relevance
05eval
RAGAS metrics with Gemini-as-judge across ten NLP test questions, rendered on a live radar chart.
Faithfulness 0.909 · overall RAGAS 0.877
RAGASGeminiPlotlyStreamlit
0.89
relevance
06infra
FastAPI RAG inference API on a t3.micro instance, Frankfurt region, with systemd auto-restart on crash or reboot and SSH key auth.
eu-central-1 · 3 REST endpoints · OpenAPI spec
AWS EC2UbuntusystemdFastAPI
0.86
relevance
07infra
The same RAG API running with two replicas under a RollingUpdate strategy — liveness and readiness probes, resource limits, zero-downtime deploys.
2 replicas · RollingUpdate · health probes
KubernetesDockerkubectl
0.83
relevance
08mlops
Three tracked experiment families — the RAG alpha search (11 runs), the QLoRA loss curve (16 steps), and the RAGAS evaluation sweep — all on a public dashboard.
11 alpha runs · public DagsHub dashboard
MLflowDagsHubPython
0.81
relevance

The fusion weight

why alpha=0.70 matters
MSc thesis finding

Sparse and dense retrieval aren't equal partners — and assuming they are costs you accuracy.

Most hybrid RAG implementations split sparse (BM25) and dense (embedding) retrieval 50/50 by default. I ran eleven systematic experiments across the fusion weight and found the true optimum sits at alpha=0.70 in favour of dense retrieval — confirmed with a paired t-test, p=0.002. That one calibration choice is worth +11.4% Recall@10 over the naive split, on a corpus of 8.84M passages.

Corpus size8.84M passages
Dense modelMicrosoft E5-base-v2
Sparse methodBM25
Experiments run11 tracked (MLflow)
Statistical testpaired t-test
Optimal alpha0.70 (p=0.002)

Background

education & experience
2026 — Present
Independent AI Engineer
Berlin, Germany — Job Seeker Visa
Built and deployed eight production AI systems covering RAG, LLM fine-tuning, multi-agent architectures, LLM evaluation, AWS EC2, Kubernetes, and MLflow tracking — all public and documented on GitHub.
Mar 2024 — Mar 2026
MSc Data Science
University of Europe for Applied Sciences — Potsdam, Germany
Thesis: hybrid RAG combining BM25 sparse retrieval with Microsoft E5-base-v2 dense embeddings. 93% Recall@10 on 8.84M MS MARCO passages. Optimal fusion weight alpha=0.70, validated with a paired t-test, p=0.002. Research informed by, and consistent with, published findings on hybrid fusion weighting in the retrieval-augmented generation literature.
currently available — Berlin

Let's build something
worth retrieving.

Actively seeking ML Engineer, NLP Engineer, and AI Engineer roles in Berlin. Job Seeker Visa holder, EU Blue Card eligible. I bring eight deployed systems and a habit of debugging production ML infrastructure until it actually works.