Prathamesh Mahalebuilding the geometry of meaning.

Bengaluru, India

• Built Hubble, a two-stage retrieval + reranking system over a 500K+ story catalog for OTT adaptation discovery (Netflix, Prime Video, SonyLIV). To bound latency and token cost, the LLM reranker never sees more than 100 candidates regardless of catalog size — embedding-based recall does the heavy lifting upstream. Cut analyst discovery time from 4h → 30min and drove a 2× increase in IP exploration.
• Designed the retrieval stack to keep the expensive LLM step at the end of the funnel: Gemini 3 Flash decomposes each brief into a tag set, Qwen (1024-d) embeddings + cosine ANN narrow 500K stories to a top-1K candidate pool, a top-100 working set is sliced by cosine score, and Gemini reranks those against the original brief for the top-10. 95% top-10 relevance on internally labeled queries from IP and licensing teams.
• Built StorySpine, a hierarchical LangGraph workflow that converts full-length stories into micro-drama adaptation plans. A supervisor agent coordinates chapter-grounded sub-agents via structured state passing — keeping each sub-agent's context window scoped to its own chapter slice while preserving narrative consistency across the full story. Two modes trade cost for quality: fast (Gemini 3.1 Flash Lite) for volume, deep (GPT-5.5 x-high) for retention-critical adaptations. Increased production throughput by 8+ adaptations/week.
• Built batch embedding infrastructure for a 5M+ story catalog on Vertex AI Batch using Qwen (1024-d) embeddings. Moved generation from synchronous online inference to scheduled distributed batch jobs, publishing vectors as shared S3 artifacts consumed by retrieval, recommendation ranking, and diversity monitoring — eliminating duplicate recomputation across teams. Reduced embedding cost by 60% while improving downstream indexing throughput.

Bengaluru, India

• Implemented SVP-CF sampling for the recommendation training pipeline — a coreset-style selection method that retained representative interactions under heavy class imbalance, cutting training data volume by 60% while preserving downstream A/B test performance.
• Refactored the Facebook ad generation pipeline for modularity, observability, and idempotency, introducing a generation cache keyed on prompt + content fingerprints. Reduced OpenAI API usage by 60–80% by eliminating redundant calls for previously-generated variants, with full request/response tracing for failure inspection.
• Re-architected the pipeline using AWS Step Functions + Metaflow DAG orchestration, decomposing sequential generation stages into parallelizable tasks with retry-aware execution and intermediate artifact persistence. Increased workflow throughput by 6× while improving observability and recovery from partial failures.
• Contributed to a LangGraph-based story-writing agent that decomposes long-form generation into chapter-scoped sub-tasks under constrained context windows — producing narrative drafts 35× faster and 10× more cost-effectively than manual workflows.

Little Rock, AR

• Improved biomedical NER under severe label imbalance using weighted cross-entropy with class-frequency-inverse weighting — lifting balanced accuracy by 4% by trading majority-class precision for minority-class recall.
• Designed a BioBERT → student-model distillation pipeline using temperature-scaled soft logits and token-level alignment losses to preserve minority-class behavior under severe class imbalance. Reduced inference cost while improving F1 and balanced accuracy by 2–5%.
• Integrated focal loss into the NER training loop to down-weight easy negatives — achieving state-of-the-art balanced accuracy on the task.
• Built 2D and 3D embedding visualization workflows for BERT-family models using projection + clustering, accelerating error analysis and surfacing minority-class confusion patterns invisible in aggregate metrics.

Pune, India

• Built an OCR + LLM document-understanding pipeline to extract dimensional rules from technical hardware-manufacturing PDFs, normalizing layout variance across vendor formats before LLM extraction — improving rule-extraction accuracy by 30%.
• Fine-tuned DistilBERT on a custom rule-extraction dataset with class-weighted training to handle imbalanced rule/non-rule sentences, achieving 98% test accuracy — chosen over a larger model to fit the inference budget required for downstream review tooling.
• Built evaluation and visualization workflows over 10+ technical PDFs, surfacing extraction-failure modes by document section to guide manual review prioritization.