Real experiments, prototypes, and proof-of-concepts from the Multimodal Minds engineering lab. Built in public. Shipped raw. No polish — just signal.
Testing CrewAI vs LangGraph for orchestrating complex, multi-step agentic workflows. Benchmarking task completion rate, token cost, and latency across 5 real business scenarios.
Combining sparse BM25 keyword retrieval with dense vector embeddings using reciprocal rank fusion. Tested on a 50k document corpus — achieved 23% better precision@5 vs pure vector search.
Exploring STT → LLM → TTS pipelines with sub-800ms end-to-end latency. Testing Deepgram, Whisper Turbo, and ElevenLabs Turbo v2.5 in combination with streaming LLM responses.
Building a lightweight data drift monitor using Evidently AI connected to an Airflow DAG. Automatically triggers model retraining when statistical drift exceeds threshold across key feature distributions.
Automating extraction from complex PDFs (tables, charts, handwriting) using GPT-4V and Claude 3.5. Compared structured output quality against traditional OCR + regex pipelines at scale.
Deploying quantized LLaMA 3 (GPTQ 4-bit) on AWS Lambda + Graviton3 for ultra cost-efficient inference. Exploring cold-start mitigation with provisioned concurrency and model caching strategies.
An agent that writes Python code, executes it in a sandboxed environment, reads the error traceback, and iteratively fixes its own mistakes. Achieved 87% task success on HumanEval-style benchmarks.
A custom evaluation framework for LLM outputs in production. Automated grading via judge LLMs + human baselines, with dashboards tracking hallucination rate, instruction-following, and output consistency over time.
Using Neo4j to model entity relationships extracted from enterprise docs, then using Cypher queries as a retrieval layer alongside vector search. Testing multi-hop reasoning on compliance and legal documents.
This is where I prototype, break, and rebuild ideas before they become products or services. Every experiment here is something I'm actively exploring in the context of real business problems — not academic exercises.
Some things will fail. Some will become part of the Multimodal Minds service stack. All of it gets documented here, openly.
Experiments are released when functional, not polished.
Every experiment includes methodology, results, and honest failure analysis.
Experiments are designed with real deployment constraints in mind.
# Multimodal Minds Labs from mm_labs import Experiment # Define the experiment exp = Experiment( name="hybrid-rag-v2", tags=["rag", "retrieval"], hypothesis="BM25 + vector fusion beats pure vector" ) # Run and log results results = exp.run() exp.log({ "precision@5": 0.847, "latency_ms": 312, "cost_per_1k": 0.0042 }) # ✅ Experiment complete print("Status: SHIPPED")
New experiments drop every few weeks. Follow along on LinkedIn or get notified when something ships.